Early prediction of childhood ADHD using prenatal and early postnatal behavioural features: evaluation across six machine-learning algorithms

Early prediction of childhood ADHD using prenatal and early postnatal behavioural features: evaluation across six machine-learning algorithms | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Early prediction of childhood ADHD using prenatal and early postnatal behavioural features: evaluation across six machine-learning algorithms Zijin Wang, Hao Wu, Ruyue Zhai, Libin Zhang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7134745/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Limited attention has been devoted to developing machine-learning models that use behavioural data for the early prediction of childhood attention-deficit/hyperactivity disorder (ADHD), particularly in the United Kingdom. Therefore, this study evaluated the predictive performance of six machine learning approaches in a cohort of 9,385 children (259 with ADHD, 9,126 controls) from the UK Millennium Cohort Study. After selecting the optimal model, we comprehensively compared the relative contributions of prenatal and postnatal (0–3 years) multi-domain features to its predictive performance. Results indicated that XGBoost achieved the highest performance on the test set (AUC = 0.881), effectively balancing the rates of false positives and false negatives. Specifically, "Conduct problems" is the most significant predictor across all models. Among postnatal features, early childhood cognitive and behavioural development represented the most influential domain, contributing approximately 51.9% SHAP value to the model. Nonetheless, other domain features (e.g. prenatal features) show non-negligible contributions. By establishing robust predictive performance, this research addresses an existing gap in machine learning-based studies of childhood ADHD within the UK context. Furthermore, as the first study to quantitatively evaluate the contribution of multiple behavioural domain features to predictive model performance in ADHD, this work provides valuable insights for future model development. ADHD Early prediction Prenatal Postnatal Machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder characterised by persistent inattention and/or hyperactivity-impulsivity [1]. Children with ADHD who are identified early and receive timely intervention and treatment have better long-term outcomes [2]. For example, those diagnosed and treated in childhood show higher self-esteem and lower levels of loneliness during adolescence compared to those diagnosed and treated in adolescence [3], as well as a reduced likelihood of delinquent and risky behaviours [4]. However, early diagnosis is highly challenging in practice. In the UK, national data indicate that although the community prevalence reaches as high as 3–5% [5], actual clinical recording rates are below 1%, with only 0.51% of children aged 5–9 years diagnosed [6]. Meanwhile, traditional diagnostic tools have low sensitivity for identifying ADHD in younger children, leading to missed diagnoses [7, 8]. Such underdiagnosis or missed diagnosis of ADHD in early childhood implies that a considerable proportion of high-risk children fail to receive timely intervention. Therefore, relying solely on formal diagnosis before intervention often means missing the optimal window of opportunity for intervention. It is urgently necessary to use risk prediction tools to identify and support high-risk ADHD children early, at symptom onset or before they reach diagnostic thresholds. On the other hand, a recent systematic meta-review of systematic reviews on ADHD demonstrated that ADHD risk factors span both the prenatal period and the early postnatal years (here broadly defined as the period after birth) and involve biological, behavioural and environmental domains [9]. In the prenatal phase, such as maternal overweight [10], smoking during pregnancy [11], and low birth weight [12] have all been linked to a higher risk of offspring ADHD; in the postnatal phase, for example, injuries [13] and second-hand smoke [14] elevate ADHD risk, whereas breastfeeding appears protective [15]. These findings suggest that simultaneously incorporating prenatal and postnatal multidomain features—and thereby integrating information across developmental periods—may potentially improve early prediction of ADHD in children. In recent years, many multivariable, multi-domain machine-learning (ML) models for ADHD prediction have proliferated [16]. Compared with traditional approaches (e.g., logistic regression), ML models not only can predict ADHD just as effectively before clinical-threshold symptom onset but also are better suited to integrating heterogeneous, multidomain data. Especially non-linear ML algorithms automatically capture higher-order and non-linear feature interactions, mitigating multicollinearity and thereby offering higher accuracy and greater practical feasibility for early ADHD prediction [17, 18]. Nonetheless, several limitations persist in current early-prediction studies for childhood ADHD. First, to our knowledge, only a handful of ML models have simultaneously considered risk factors from both prenatal and postnatal domains [19–21]. Even these few studies include only limited antenatal variables and do not evaluate the relative contribution of prenatal versus postnatal exposures to model performance. Moreover, among these studies, only one investigation has attempted to predict ADHD truly early in childhood, leveraging features measured before age 5 to forecast ADHD at 5–6 years [19]. In contrast, the remaining studies examine broader age ranges from childhood to adolescence. Moreover, UK-based ML studies that specifically target paediatric ADHD are exceedingly scarce, with existing studies primarily focusing on predicting ADHD at ages 6–7 based on risk factors identified at ages 4–5 [22]. To date, no research has explored even earlier prediction within UK cohorts (e.g. using factors measured at or before age three). Although a recent preprint attempted to develop an early ADHD prediction model for British children [23], approximately 98% of healthy samples in the initial dataset were removed to balance ADHD and healthy samples, potentially inducing a prior-probability shift and somewhat inflating model performance in real-world cohorts. Given the limitations of previous studies described above, this study based on the UK Millennium Cohort Study (MCS) to predict ADHD outcomes at early school age (5–7 years) using information collected before age three. Four objectives are pursued: (1) after combining early postnatal features (0–3 years) with a broad set of maternal prenatal features, we will compare six machine-learning algorithms (Random Forest, Support Vector Machine, Light Gradient Boosting Machine, K-Nearest Neighbors, Extreme Gradient Boosting, and Multilayer Perceptron) and select the optimal model, thereby addressing the gap in UK research on predicting ADHD using information available prior to age three; (2) enhance model generalisability by applying sample weights for ethnic-minority and low-income groups; (3) employ SHAP analysis to provide an interpretable account of the key predictors identified by the optimal model; and (4) use SHAP values to assess the relative contributions of different domains to overall model performance. Methods Participants and design The dataset used in this study is the Millennium Cohort Study (MCS). The MCS aims to track the development of UK children born in 2000 or 2001 [24]. Data collection methods include questionnaires and interviews, which continuously gather information on the health, cognitive development, educational achievement, and socioeconomic background of cohort members. However, the sample structure of the MCS has limitations, including geographically uneven distribution and disproportionate representation of ethnic minorities and families from disadvantaged areas. Therefore, clustering, stratification, and weighting methods are required to post hoc correct model parameters and balance population representativeness [24]. All cohort members participating in this study provided informed consent signed by themselves or by their caregivers acting as proxies [25]. Definition of outcome and Sample derivation The present study selected early ADHD diagnosis data as the outcome, at age 5 (sweep 3) and age 7 (sweep 4). To better define the composition of the early ADHD diagnostic group, participants who were not diagnosed in early childhood but received an ADHD diagnosis at age 11 (sweep 5) or age 14 (sweep 6) (i.e., late childhood and early adolescence) were excluded, thus controlling for their potential confounding effect as early “healthy controls” in the healthy group. The early and late ADHD diagnostic data were obtained from four datasets within the MCS: mcs3_parent_cm_interview and mcs4_parent_cm_interview, as well as mcs5_parent_cm_interview and mcs6_parent_cm_interview, respectively. The variables CPADHD00 and DPADHD00 (reported by mothers) constituted the early ADHD diagnosis, whereas the variables EPADHD00 and FPADHD00 constituted the late ADHD diagnosis (i.e., whether diagnosed with ADHD). Given missing data at these four time points, the following rules were applied to select participants for early/late ADHD diagnostic groups: 1) For early ADHD diagnosis, participants who reported a confirmed ADHD diagnosis at either age 5 or age 7 (i.e., CPADHD00 or DPADHD00 = 1) were classified into the ADHD diagnostic group. Participants who reported no ADHD diagnosis at both ages 5 and 7 (i.e., CPADHD00 and DPADHD00 = 2) were classified into the healthy group. Participants who did not meet either criterion were marked as missing. 2) For late ADHD diagnosis, participants who reported a confirmed ADHD diagnosis at either age 11 or age 14 (i.e., EPADHD00 or FPADHD00 = 1) were classified into the ADHD diagnostic group. Participants who reported no ADHD diagnosis at both ages 11 and 14 (i.e., EPADHD00 and FPADHD00 = 2) were classified into the healthy group. Participants who did not meet either criterion were marked as missing. The samples for the early/late diagnostic groups identified based on the above rules underwent further selection procedures, as detailed in Table 1. Sweep 3 initially included 15,373 cohort members, of whom 9,385 remained after screening. Among these, 9,126 participants were classified into the healthy control group without an ADHD diagnosis, and 259 (2.76%) were classified into the ADHD diagnostic group. Table 1. Sample selection pipeline. Data collection The time range of feature selection in this study spans from the mother's early pregnancy to the child's third birthday. Apart from demographic features, the features in this study were divided into two main domains based on the child’s birth: prenatal features and early postnatal features (i.e. early childhood period, ages 0-3). Features with more than 10% missing values were excluded. Additionally, paternal ADHD-related features were entirely excluded due to substantial missing data in the MCS; therefore, this study only included mother-reported features. Ultimately, a total of 45 features were included in this study. Four demographic features were included: Gender, Race, Household income quintiles, and Mothers academic qualification. Fifteen prenatal features were included: Premature, Low weight, Prenatal depression, Prenatal eclampsia, Prenatal epilepsy, Prenatal Suspected Slow Growth, Multiple pregnancy, Threatened miscarriage, Prenatal asthma, Prenatal persistent vomiting, Prenatal non-trivial infections, Prenatal smoking, Prenatal drinking, Age group, and Prepregnancy BMI. Twenty-six early postnatal features were included and further divided into three sub- domains: (1) Early maternal parenting style: Mother-child Conflicts, Mother-child positive relationship, Breastfeeding, Postpartum depression, Regular bedtimes, Regular eat, Single parent; (2) Child environmental exposure: Media, Injury, Secondhand smoke; (3) Early child cognitive and behavioural development: Suspected social delay, Suspected fine motor delay, Suspected gross motor delay, Mood, Approach, Adaptability, Regularity, Cry, Naming vocabulary, School readiness composit, Independence-Self Regulation, Emotional-Dysregulation, Emotional Symptoms, Conduct problems, Peer problems, Prosocial Behaviour. For more details, see Table A1in the Appendix. Given that maternal age during pregnancy showed a nonlinear relationship with offspring ADHD [26], Age group was one-hot encoded similarly to Race, using "Pregnant at the age of 12-19" as the baseline level. After one-hot encoding, the total number of features increased to 53. Data preprocessing For features with less than 10% missing data, nonlinear multiple imputation was performed using miceforest 0.6.3 [27] in Python 3.12.3. Missing values for the Sweep 4 weighting variable (i.e., DOVWT2) were set to 1, indicating that they did not participate in weight adjustment [28]. The Dataset was randomly divided using train_test_split, with a fixed random seed of 42 and a 7:3 ratio between the training and test sets. Statistical analysis This study evaluated six nonlinear ML algorithms, including tree-based models Random Forest (RF) [29], Extreme Gradient Boosting (XGBoost) [30], Light Gradient Boosting Machine (LightGBM) [31]; and non-tree-based models Support Vector Machine (SVM) [32], K-Nearest Neighbors (KNN) [33], Multilayer Perceptron (MLP) [34]. The rationale for selecting these algorithms was that the dataset had a relatively large sample size and contained highly nonlinear features, including binary, categorical, and continuous variables. RF, XGBoost, KNN, LightGBM, and MLP can directly handle high-domain nonlinear data, while the SVM utilises a Radial Basis Function (RBF) kernel to handle such nonlinear relationships. ML analyses were implemented using Python 3.7 and Scikit-learn 0.24.1. The code is freely available at GitHub: https://github.com/Anankysus/Early-ADHD-ML/tree/main. Following the User Guide to Analysing MCS Data using Stata [35], the primary sampling unit variable SPTN00 and the overall sample weight DOVWT2 (national sampling weights based on Sweep 4) were selected for model development and evaluation, thus accounting for clustering correlations and sampling probability differences. In this study, the ratio of ADHD diagnostic cases to healthy controls was 1:35.3, reflecting extreme class imbalance due to the predominance of the healthy group. Therefore, an additional validation set was not created; instead, cross-validation was used. To mitigate the issue of ML models excessively focusing on the majority (healthy) class due to this imbalance, three balancing strategies were employed: Class Weights [36], Random Undersampler [37], and threshold tuning using Youden’s J index [38]. Specifically, the Random Undersampler ratio was set to 0.2 (random seed = 42), resulting in a healthy-to-ADHD ratio of 905:181 in the training set. Empirical testing showed that a ratio of 0.2 not only yielded optimal model performance but also outperformed oversampling methods such as SMOTE and SMOTE-ENN [39]. To minimise the potential impact of redundant features on ML models, Recursive Feature Elimination (RFE) [40] was employed for feature selection in each of the six models to determine the optimal subset of features. RFE and model training were embedded in the same pipeline, where both feature selection and hyperparameter tuning (GridSearchCV) were conducted simultaneously within the same cross-validation framework. All ML hyperparameter settings are detailed in Table A2 in the Appendix. Within this pipeline, to further avoid information leakage among samples within the primary sampling unit (PSU) during hyperparameter tuning, Group K-Fold cross-validation (with MCS’s cluster identifier SPTN00 as the grouping factor) was conducted on the training set. Specifically, a 10-fold group-wise cross-validation was used, with 9-folds for training and 1-fold for validation in each iteration. The random seed was fixed at 42 for all models. After completing the training phase, each base classifier’s raw prediction scores were calibrated using 10-fold sigmoid calibration to correct model confidence biases caused by extreme class imbalance. To ensure unbiased estimation of all performance metrics according to the target population, complex survey sampling weights from MCS were incorporated, along with class weights, to form composite weights integrated into each ML model. Performance metrics were calculated using three averaging approaches: macro-averaging, micro-averaging, and weighted averaging. Macro-averaging calculates metrics independently per class and then averages these; micro-averaging merges predictions and true labels across classes before calculating overall metrics; weighted averaging computes metrics per class and then averages these metrics according to each class’s composite weight in the test set. Detailed evaluation metrics included Accuracy, Precision, Specificity, Recall, and F1-score. Receiver Operating Characteristic (ROC) curves and corresponding Area Under the Curve (AUC) were used to measure the overall model discrimination between positive and negative classes across all thresholds. Additionally, Precision–Recall (PR) curves were plotted, which emphasise the trade-off between recall and precision for the minority class (ADHD group) in highly imbalanced datasets, complementing the ROC curves that may lack sensitivity for minority-class performance. Confusion matrices were generated for each ML model to illustrate the distribution of true positives, false positives, true negatives, and false negatives, aiding identification of specific misclassification patterns. Composite weights were also incorporated into ROC, confusion matrix, and PR curve analyses to mitigate biases resulting from imbalanced classes and sampling design. Feature importance was reported for the model with the best test performance, reflecting each variable’s overall contribution to model performance (taking composite weights into account). However, feature importance only indicates the critical features influencing model discrimination, without clarifying whether their effects are positive or negative at various values. Therefore, SHAPley Additive exPlanations (SHAP) [41, 42] were employed for detailed analysis of the best-performing model. SHAP decomposes the model’s output into additive feature contributions, enabling both a global feature ranking (by cumulative contributions) and local analysis of each feature’s specific values, enhancing model transparency and reliability [41]. The global SHAP feature influences were presented as violin plots, using replicated samples to simulate composite weights, while local feature-value impacts based on SHAP values were illustrated using boxplots Lastly, cross- domain group comparisons of contributions to the model were performed based on SHAP values. Results Clinical characteristics Demographic features of this study are shown in Table 2. Among the total sample of 9385 cohort members, there were 4769 girls (50.8%) and 4616 boys (49.2%). Ethnicity was predominantly White (85.0%). The distribution of participants across Household income quintiles was relatively even, with the Lowest quintile having the smallest proportion (15.17%) and the other four quintiles ranging between 20% and 21%. Regarding Mothers academic qualification, the largest group was O level/GCSE grades A-C (30.5%), while the smallest proportion was Higher degree (3.8%). Within specific groups, gender was evenly distributed in the healthy group but severely imbalanced in the ADHD diagnostic group, with a female-to-male ratio of approximately 1:4. White ethnicity dominated both healthy and ADHD groups, slightly more so in the ADHD diagnostic group. The distribution of Household income weighted quintiles was relatively balanced in the healthy group; however, the lowest-income group dominated (39.0%) in the ADHD diagnostic group. Additionally, mothers with higher academic qualifications were generally less represented in the ADHD diagnostic group compared to the healthy group. Table 2. Participant demographic features. Sample characteristic No. (%) ADHD diagnosed Healthy Total No. of samples 259 9126 9385 Gender Girl 50(19.3) 4719(51.7) 4769(50.8) Boy 209(80.7) 4407(48.3) 4616(49.2) Missing 0 0 0 Race White 231(89.2) 7764(84.9) 7977(85.0) Mixed 8(3.1) 243(2.7) 251(2.7) Indian 5(1.9) 227(2.5) 232(2.5) Pakistani and Bangladeshi 8(3.1) 545(6.0) 553(5.9) Black or Black British 5(1.9) 247(2.7) 252(2.7) Other Ethnic group 2(0.8) 115(1.3) 117(2.7) Missing 0 3(0.0003) 3(0.0003) Household income quintiles Lowest quintile 101(39.0) 1568(17.2) 1669(17.8) Second quintile 69(26.6) 1839(20.2) 1908(20.3) Third quintile 36(13.9) 1865(20.4) 1901(20.3) Fourth quintile 30(11.6) 1921(21.0) 1951(20.8) Highest quintile 23(8.9) 1923(21.1) 1946(20.7) Missing 0 10(0.1) 10(0.1) Mothers academic qualification Higher degree 5(1.9) 350(3.8) 355(3.8) First degree 24(9.3) 1317(14.4) 1341(14.3) Diplomas in higher education 19(7.3) 849(9.3) 868(9.2) A/AS/S levels 24(9.3) 867(9.5) 891(9.5) O level/GCSE grades A-C 75(29.0) 2785(30.5) 2860(30.5) GCSE grades D-G 17(6.6) 842(9.2) 859(9.2) Other academic qualifications 6(2.3) 233(2.6) 239(2.5) None of these qualifications 52(20.1) 1303(14.5) 1355(14.4) Missing 37(14.3) 580(6.4) 617(6.5) Predictive performance The optimal hyperparameters for the six ML models are presented in Table A3 in the Appendix. After applying RFE, each model retained 22 variables. Table 3 presents the weighted accuracy scores for the training set and testing set of each ML model. On the test set, XGBoost achieved the highest weighted accuracy (0.8126), while the weighted accuracies of the other five models ranged from 0.7 to 0.8, with LightGBM having the lowest weighted accuracy (0.7087). Comparing weighted accuracy between the training and testing sets, RF and LightGBM exhibited potential overfitting, with a gap greater than 10%. Table 3. Weighted accuracy Scores of Training Set and Testing Set for six machine learning algorithms Algorithms ADHD diagnosed group vs Healthy group Training set Test set Random Forest 0.9346 0.7918 SVM 0.7882 0.7954 LightGBM 0.9659 0.7087 KNN 0.7716 0.7763 XGBoost 0.8453 0.8126 MLP 0.7845 0.7945 Table 4 presents the evaluation metrics for the six models based on macro-, micro-, and weighted-averaging approaches, and Fig. 1 displays the weighted confusion matrix results for each model. After accounting for sample frequency and sampling weights, weighted F1 was adopted as the primary evaluation metric, supplemented by weighted precision, weighted recall, macro F1, and diagnostic recall rate. XGBoost achieved a weighted F1 score of 0.813, a weighted precision of 0.820, a weighted recall of 0.813, and a macro F1 score of 0.585 on the test set. Its confusion matrix showed a diagnostic recall rate of 76.3%. Considering these comprehensive metrics, XGBoost outperformed all other models. Combined with the highest weighted accuracy, XGBoost was selected as the preferred model due to its superior classification performance and robustness to extreme class imbalance. Its optimal threshold was 0.1300, with a Youden’s J index of 0.611. Table 4 . Six machine learning algorithms performance for ADHD diagnosed Algorithms Specificity Precision Recall F1 score Random Forest macro 0.9270 0.5963 0.7904 0.6336 micro 0.9270 0.9194 0.9194 0.9194 weighted 0.9203 0.8180 0.7918 0.7901 SVM macro 0.8327 0.5540 0.8010 0.5534 micro 0.8327 0.8310 0.8310 0.8310 weighted 0.8128 0.7975 0.7954 0.7956 LightGBM macro 0.9383 0.5884 0.7255 0.6200 micro 0.9383 0.9265 0.9265 0.9265 weighted 0.9367 0.7740 0.7087 0.6977 KNN macro 0.7779 0.5398 0.7672 0.5150 micro 0.7779 0.7773 0.7773 0.7773 weighted 0.7651 0.7766 0.7763 0.7764 XGBoost macro 0.8795 0.5667 0.7923 0.5846 micro 0.8795 0.8746 0.8746 0.8746 weighted 0.8704 0.8201 0.8125 0.8126 MLP macro 0.7893 0.5458 0.7985 0.5275 micro 0.7893 0.7898 0.7898 0.7898 weighted 0.7708 0.7943 0.7945 0.7944 Aside from LightGBM, the weighted accuracy of the remaining four models did not differ significantly from that of XGBoost (Holm-adjusted p-value > 0.05) (See Table A4 in the Appendix). However, confusion matrices indicated limitations and shortcomings in suppressing false positives and false negatives in each model. Among these, SVM had a minority-class recall rate of 78.0%, demonstrating slightly higher sensitivity and thus a lower risk of false negatives compared to XGBoost, but at the cost of approximately a 6% increase in false-positive rate. Nevertheless, due to its recall rate of over 80% for true positives, SVM was identified as the second-best model. ROC and PR curves further validated the robust predictive performance of XGBoost in the classification task (see Fig. 2). XGBoost achieved the highest AUC (0.881) and the second-highest Average Precision (AP = 0.900) on the test set, with both curves maintaining high specificity and precision at high recall levels. In contrast, although RF achieved the highest AP (0.908) and second-highest AUC (0.879) on the PR curve, these values were inflated by its exceptionally high recall of the healthy group (92.03%), whereas its diagnostic recall rate was the second lowest (68.11%) among the six models (see Fig. 1A). Although SVM, LightGBM, KNN, and MLP were not selected as preferred models, they all obtained AUC values greater than 0.84 and AP values exceeding 0.86, demonstrating similarly high accuracy for overall classification. Feature importance Fig. 3A shows the ranking of feature importance for the optimal XGBoost model. Early child cognitive and behavioural development features dominated numerically, comprising 10 features (approximately 45.5%): Conduct problems, Peer problems, Prosocial behaviour, School readiness assessment, Emotional-Dysregulation, Naming vocabulary, Cry, Regularity, Adaptability, and Approach. There were 5 prenatal features: Age_group_2 (Pregnant at the age of 20-29), Prenatal smoking, Age_group_3 (Pregnant at the age of 30-39), Prepregnancy BMI, and Prenatal drinking; Four early maternal parenting style features: Breastfeeding, Single parent, Regular bedtimes, and Regular eat; Two demographic features: Gender and Mothers academic qualification; and 1 children environmental exposure feature: Secondhand smoke. Given that tree-based and non-tree-based models differed in terms of feature selection and ranking of feature importance, and given that the performance differences between models were not significant in this study (See Fig. A1 in the Appendix), the present study additionally examined the feature composition of non-tree-based models, using SVM as the best-performing non-tree-based model (see Fig. 3B). Among the 15 features positively contributing to the SVM model, only 3 belonged to early child cognitive and behavioural development: Conduct problems, Emotional-Dysregulation, and Suspected social delay. In contrast, prenatal features accounted for the largest proportion (approximately 45.3%), including Threatened miscarriage, Prenatal non-trivial infections, Premature, Prenatal smoking, Low weight, and Age_group_3 (Pregnant at the age of 30-39). Additionally, there were 3 demographic features: Race_4 (Pakistani and Bangladeshi), Gender, and Race_5 (Black or Black British); 1 environmental exposure feature: Secondhand smoke; and 2 early maternal parenting style features: Breastfeeding and Postpartum depression. Overall, prenatal factors accounted for approximately 22.7% of the selected features in the XGBoost model. In contrast, prenatal factors accounted for approximately 45.2% of the effective features in SVM, the representative non-tree-based model. Thus, the proportion of prenatal features retained in SVM exceeded that of XGBoost. To further explore the feature selection preferences of these two models regarding prenatal and postnatal domains, two sensitivity analyses were conducted: a linear correlation analysis between prenatal and early postnatal feature groups and a permutation-based sensitivity analysis using 10-fold cross-validation (see Table A5 & Table A6 in the Appendix). The linear correlation analysis revealed minimal linear redundancy and high independence between these two feature groups, with a mean absolute correlation coefficient (|ρ|) of 0.0301 and a 75th percentile of less than 0.037, indicating that very few pairs exceeded a correlation of 0.10. The permutation sensitivity analysis showed that for XGBoost, the average AUC decreased by only 0.0129 after permutation of prenatal features, whereas it decreased by 0.2440 after permutation of postnatal features. In contrast, for SVM, average AUC decreased by 0.0274 after permutation of prenatal features—approximately 2.1 times greater than for XGBoost—and decreased by 0.1756 after permutation of postnatal features, approximately 0.72 times that of XGBoost. SHAP Analysis of XGBoost This study presents the SHAP value distributions of the 10 features with the highest contributions to the model output (see Fig. 4), collectively explaining 91% of the model’s predictive output, thus dominating overall prediction. Features are sorted from top to bottom based on mean absolute SHAP values. The horizontal axis represents the SHAP value, where positive SHAP values indicate a feature pushing predictions toward the ADHD diagnostic group (positive class), and negative SHAP values indicate predictions toward the healthy group (negative class). Each violin plot’s colour represents the original feature values, with red indicating higher feature values and blue indicating lower feature values. In general, for continuous features (Conduct problems, Peer problems, Emotional-Dysregulation), higher feature values were associated with increased SHAP values, driving predictions toward the ADHD diagnostic group. Among binary features, lower Gender values (i.e., male) predicted the diagnostic group, whereas higher Age_group_3 values (Yes) predicted the healthy group. High values for other binary features such as Breastfeeding (No), Age_group_2 (Yes), Single parent (Yes), and Prenatal smoking (Yes) also predicted classification in the ADHD diagnostic group. Finally, the ordinal feature Mothers academic qualification showed that higher academic qualification predicted classification in the diagnostic group. Based on the global SHAP summary plot, this study further examined local feature-value impacts, illustrating how specific values within each of the top 10 features influenced diagnostic predictions (see Fig.5). Specifically, the feature trends for binary variables (i.e., Gender, Age_group_3, Breastfeeding, Age_group_2, Single parent, and Prenatal smoking) aligned with the global SHAP results. For continuous variables, higher scores on Conduct problems (>3 points, range 0–10), Peer problems (>2 points, range 0–10), and Emotional-Dysregulation (>1.57 points, range 1–3) were more predictive of ADHD diagnosis. Regarding the ordinal variable, Mothers academic qualification, qualifications higher than O level/GCSE grades A–C were more predictive of ADHD diagnosis. This result contradicts findings from multinational cohorts and previous studies based on the UK MCS dataset [43]. To confirm that this reversal was not a statistical artefact, two sensitivity analyses were conducted (see Table A7 & Fig. A2 in the Appendix). Univariate logistic regression indicated that each decrease in educational qualification level corresponded to an OR = 1.14 (p < 0.001), consistent with prior studies. Additionally, removing sampling weights from the SHAP analysis did not alter the direction of this feature, indicating that its association was not influenced by sampling weights. Finally, using cumulative absolute SHAP values, the contributions of the ten most important features across four predefined domains were quantified and compared, together these domains account for 100 % of the total contribution (Fig. 6). The domain of early child cognitive and behavioural development contributed over half (51.9%), whereas domain of demographic features, prenatal features, and maternal early parenting style contributed comparably, accounting for 19.8%, 14.7%, and 13.6% respectively. Discussion This study predicts early childhood ADHD at ages 5–7 based on features collected before age 3, combining prenatal and postnatal multidomain features. Among the six evaluated models, XGBoost exhibited the best overall performance, achieving an AUC of 0.881. This places the model near excellent performance among ADHD diagnostic models (where reported AUCs range between 0.50 and 0.96) and ranks highly even among prediction models relying solely on behavioural data [ 11 ]. Moreover, the confusion matrix for XGBoost indicates that the model effectively identifies ADHD cases while minimising false negatives. Regarding feature composition, considering that the five-item Hyperactivity/Inattention (H/I) subscale from the SDQ is similar with DSM-IV/V ADHD diagnostic criteria, including this subscale in the model would likely result in symptom-overlap effects, inflating internal test performance and reducing the model’s capacity to detect other early risk signals. To emphasize genuine predictive value rather than simply "earlier administration of a diagnostic scale," this study excluded the H/I subscale, despite the potential slight decrease in sensitivity. Nevertheless, all models in this study maintained good-to-excellent predictive capability. The feature importance results revealed differences between XGBoost and SVM in their selection of prenatal and postnatal features, suggesting that SVM may show a stronger preference for prenatal factors. Further sensitivity analyses demonstrated that, even under conditions of weak correlations between prenatal and postnatal feature groups, XGBoost tended to prioritize only a limited number of high-gain postnatal features [ 44 ]. In contrast, SVM’s RBF kernel distance mechanism enabled it to incorporate weaker but stable signals from both prenatal and postnatal features [ 45 ]. These results further support the idea that SVM favours prenatal features, whereas XGBoost focuses predominantly on a smaller subset of postnatal features, providing the highest gain. Such findings have implications for future development of early ADHD prediction tools: the choice of algorithm itself may influence the preference for predictive features across different time frames and multidomain domains. Further research is needed to validate these algorithm-specific preferences in other datasets. SHAP Analysis of XGBoost A major strength of this study is the SHAP analysis performed on the XGBoost model, which provided a comprehensive evaluation and interpretation of how specific feature values affect ADHD prediction. Conduct problems from the Strengths and Difficulties Questionnaire (SDQ) [ 46 ] contributed the highest mean marginal impact on the model output. Notably, Conduct problems ranked distinctly first in terms of feature importance, a finding consistently observed across other models as well (see Fig. A1 in the Appendix). Based on these findings, this study suggests that early childhood Conduct problems may serve as one of the most critical behavioural indicators for predicting subsequent ADHD. Existing literature supports this hypothesis: previous SDQ-based studies have found that children with ADHD consistently exhibit elevated Conduct problems [ 47 ], and multivariate analysis of SDQ subscales at age 3 further demonstrated that, apart from the H/I subscale, Conduct problems is the only item independently predictive of ADHD diagnosis two years later [ 48 ]. Peer problems from the SDQ [ 46 ] and Emotional-Dysregulation from the Child Social Behaviour Questionnaire (CSBQ) [ 49 , 50 ] also significantly contributed to predicting subsequent ADHD, though their impacts were not as pronounced as Conduct problems. Previous longitudinal cohort studies have reported that children with ADHD score significantly higher on Peer problems compared to healthy populations, a trend that persists longitudinally [ 51 ]. Emotional-Dysregulation is similarly prevalent among children with ADHD, with meta-analyses indicating it may affect approximately 25–45% of this population [ 52 ]. Interestingly, although Emotional Symptoms from the SDQ [ 46 ] was also considered as a candidate emotional feature, it proved consistently less crucial than Emotional-Dysregulation across all models, both in feature importance rankings and SHAP analyses. This suggests Emotional-Dysregulation is more closely aligned with core ADHD processes. Prior findings also support this interpretation, indicating that Emotional-Dysregulation acts as a mediator between ADHD and Emotional Symptoms. After controlling for this mediating effect, the direct influence of ADHD on Emotional Symptoms is not significant [ 53 ]. Further, local feature-value impacts based on SHAP values indicated that SHAP contributions for Conduct problems, Peer problems, and Emotional-Dysregulation tended to plateau at lower scores, stabilising marginal SHAP values. For Conduct problems and Peer problems, the inflexion points in SHAP values were below the "Borderline" thresholds (approximately top 20% in the normative data) established by SDQ for UK children aged 2–4 years [ 54 ], suggesting that the model was able to identify symptom levels below official clinical cut-offs as potential early-risk signals. For Emotional-Dysregulation, despite the absence of a formal cut-off [ 55 ], the SHAP inflexion point (1.862) was below the sample mean in this study, with approximately 75% of children scoring above this inflection point. In fact, these apparently "low-score, high-risk" cases are not misclassifications. XGBoost models do not rely on predefined thresholds or percentile ranks; instead, they identify feature splits with the maximum information gain directly from the training data [ 31 ]. Thus, samples below established scale thresholds or percentiles may trigger secondary splits in decision trees, generating positive SHAP contributions and predicting high-risk status [ 56 ]. Furthermore, approximately 21% and 29% of children in this study scored ≥ 3 for Conduct problems and ≥ 2 for Peer problems, respectively. Although these scores do not strictly surpass the borderline threshold (20%), they closely align with the theoretical proportion for the "Borderline" range on the SDQ, indicating these children represent a borderline risk zone warranting greater practical attention. SHAP analysis results for several binary variables were consistent with previous studies, indicating that being male and not receiving breastfeeding predicted children's subsequent risk of ADHD, a conclusion supported by a recent systematic meta-review of systematic reviews on ADHD [ 9 ]; similarly, single parent and prenatal smoking have similarly been shown by previous systematic reviews and meta-analyses to predict ADHD [ 57 , 58 ]. Compared with the baseline category "Pregnant at the age of 12–19," pregnancy at the age of 30–39 was more predictive of the healthy group, consistent with previous meta-analysis findings [ 26 ]. In contrast, pregnancy at the age of 20–29 tended to predict the ADHD group, contradicting the commonly reported conclusion in the literature that pregnancy before age 20 carries the highest risk. This discrepancy may result from RFE excluding the "Pregnant at the age of 40–49" category, causing both the 12–19 and 40–49 age categories to be coded as zero in the dummy-variable matrix, thus forming a composite baseline. Although RFE improved model performance, it reduced the interpretability of SHAP values for the maternal age variable. Therefore, SHAP results for pregnancy at the age of 20–29 should be interpreted cautiously. Regarding Mothers academic qualification, since sampling design factors were excluded, it can be inferred that after entering multiple features simultaneously into the model, the risk associated with lower educational attainment was absorbed by other multidomain ADHD-related variables, thereby weakening the marginal effect of education itself [ 59 ]. Additionally, families with higher educational qualifications typically have better access to healthcare and diagnostic awareness, resulting in more children from these families receiving diagnoses [ 60 ]. Thus, the positive SHAP value associated with education is more likely reflective of diagnostic behaviour differences and covariate collinearity, rather than indicating that higher education itself increases ADHD risk. Finally, this study systematically integrated the domains corresponding to the top ten features, which represent the majority of model contributions, and conducted cross-domain comparisons. To our knowledge, this is the first ADHD prediction study conducted within a behavioural framework to perform cross-domain group comparisons. Results indicated that the domain of early child cognitive and behavioural features contributed most significantly to the model. First, the features included in this domain were temporally closer to the diagnostic assessments at ages 5–7, allowing the model to better capture the most discriminative symptomatic manifestations during this period [ 59 , 61 ]. This temporal proximity may have partly diluted the impact of demographic-domain features—specifically mother’s academic qualification and household income quintiles—in the ML model; at the same time, it may also have absorbed some of the explanatory power of the prenatal-domain features, thereby reducing that domain’s independent contribution. Second, previous studies have demonstrated that Conduct problems, Peer problems, and Emotional-Dysregulation serve as external indicators of restricted executive functioning in children and are significantly associated with subsequent ADHD symptoms [ 62 – 64 ], whereas executive-function deficits are widely regarded as one of the core neurocognitive impairments in ADHD [ 65 ]. Therefore, this study hypothesises that early potential deficits in executive functioning may amplify ADHD risk through the expression of these specific features. However, this study did not include executive functioning features, and thus, the underlying mechanisms require further objective testing for validation. Nevertheless, prenatal factors and early maternal parenting style still jointly contributed 28.3% to the model's explanatory power, suggesting they carry independent information not covered by early child cognitive and behavioural development or demographic features. As these risk factors can be modified through public health interventions (e.g., smoking cessation during pregnancy, promotion of breastfeeding) and family interventions (e.g., parental behavioural management training) before ADHD symptoms emerge, they remain irreplaceably valuable for early screening and risk management. limitations The present study has several limitations. First, due to computational constraints, this study did not exhaust all possible algorithmic combinations to achieve optimal test-set performance—for example, simultaneously incorporating undersampling, RFE, hyperparameter tuning, calibration, and threshold optimisation within a single pipeline to perform global optimisation. Such an approach would exponentially increase computational time due to multiple nested exploratory processes. Second, given the scarcity of ADHD cases among children, this study was unable to perform stratified analyses based on specific demographic variables. Instead, demographic variables were collectively incorporated into the ML models. Particularly regarding the Gender variable, the demographic descriptive analysis indicated a significant gender imbalance within the ADHD diagnostic group, potentially making Gender the most crucial feature in multiple models. Third, to address training difficulties caused by extreme sample imbalance, RFE-based feature selection was conducted before model training. Although this improved model performance to some extent, it exacerbated the imbalance in the number of selected features across domains. For instance, the environmental exposure domain retained only one feature in the optimal XGBoost model, hindering meaningful cross-domain comparisons. Fourth, this study utilised only binary diagnostic labels from the MCS dataset as supervision signals, thereby limiting further exploration of specific core ADHD symptoms and the influence of corresponding clinical assessment scores within the machine learning framework. Although the MCS dataset contains SDQ Hyperactivity/Inattention (H/I) items with detailed scale scores, allowing for potential mapping onto core ADHD symptoms, the SDQ scale itself is not a clinical diagnostic instrument, and its results lack diagnostic authority for ADHD. Conclusion This study developed an early childhood ADHD prediction model using the XGBoost machine learning algorithm. Its predictive performance was not only favourable compared to existing early childhood ADHD prediction models but also demonstrated the capacity to predict ADHD diagnosis at ages 5–7 using behavioural-level data collected by age 3. More importantly, these findings are generalizable to the entire UK population. Although early child cognitive and behavioural features played a critical role in ADHD prediction, maternal prenatal factors and other postnatal domains should not be overlooked. In conclusion, this study offers valuable insights for future ADHD machine learning research, underscoring the importance of comprehensively considering the contributions of multidomain features to ADHD prediction. Declarations Competing interests : The authors declare no competing interests Ethical approval : All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Consent to participate: All cohort members participating in this study provided informed consent signed by themselves or by their caregivers acting as proxies Author Contribution ZW: Contributed to the conception or design of the study, data analysis and interpretation, manuscript writing, and revision and review of the manuscript.HW: Contributed to partial manuscript writing, revision and review of the manuscript.RZ: Contributed to revision and review of the manuscript.LZ: Contributed to the conception or design of the study, revision and review of the manuscript. Data Availability The data used in this study were derived from the UK Millennium Cohort Study (MCS), which is publicly available for research purposes. The data on maternal prenatal and child postnatal variables from birth to age 7 that support the findings of this study have been deposited in the UK Data Service repository at: https://ukdataservice.ac.uk.The datasets involved in this study are classified as safeguarded level by the UK Data Service. Due to the sensitive nature of the data, raw data and individual-level identifiers cannot be publicly shared. Researchers wishing to access the MCS dataset must register with the UK Data Service and agree to the terms and conditions of data use.The analysis code has been made publicly available on GitHub: https://github.com/Anankysus/Early-ADHD-ML/tree/main. Additionally, detailed explanations of the variables/features used in the study are provided in the supplementary materials, to facilitate replication of the study. References American Psychiatric Association (2022) Diagnostic and statistical manual of mental disorders. Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR) 5:. https://doi.org/10.1176/appi.books.9780890425787 Brites C, Brites HD, de Almeida RP et al (2023) Early Diagnosis on ADHD. Is It Possible? Psychol 14:359–370. https://doi.org/10.4236/psych.2023.143021 Zahmacioglu O, Kilic E (2017) Early diagnosis and treatment of ADHD are important for a secure transition to adolescence. Anatol J Psychiatry 18:1. https://doi.org/10.5455/apd.220054 Boland H, DiSalvo M, Fried R et al (2020) A literature review and meta-analysis on the effects of ADHD medications on functional outcomes. J Psychiatr Res 123:21–30. https://doi.org/10.1016/j.jpsychires.2020.01.006 Ford T, Goodman R, Meltzer H (2003) The British Child and Adolescent Mental Health Survey 1999: the Prevalence of DSM-IV Disorders. J Am Acad Child Adolesc Psychiatry 42:1203–1211. https://doi.org/10.1097/00004583-200310000-00011 Prasad V, West J, Kendrick D, Sayal K (2019) Attention-deficit/hyperactivity disorder: variation by socioeconomic deprivation. Arch Dis Child 104:802–805. https://doi.org/10.1136/archdischild-2017-314470 Øvergaard KR, Oerbeck B, Friis S et al (2023) Early and repeated screening detects children with persistent attention-deficit/hyperactivity disorder. Eur Child Adolesc Psychiatry. https://doi.org/10.1007/s00787-023-02284-8 Øvergaard KR, Oerbeck B, Friis S et al (2018) Attention-Deficit/Hyperactivity Disorder in Preschoolers: The Accuracy of a Short Screener. J Am Acad Child Adolesc Psychiatry 57:428–435. https://doi.org/10.1016/j.jaac.2018.03.008 Chaulagain A, Lyhmann I, Halmøy A et al (2023) A systematic meta-review of systematic reviews on attention deficit hyperactivity disorder. Eur Psychiatry 66. https://doi.org/10.1192/j.eurpsy.2023.2451 Li L, Lagerberg T, Chang Z et al (2020) Maternal pre-pregnancy overweight/obesity and the risk of attention-deficit/hyperactivity disorder in offspring: a systematic review, meta-analysis and quasi-experimental family-based study. Int J Epidemiol 49:857–875. https://doi.org/10.1093/ije/dyaa040 Dong T, Hu W, Zhou X et al (2018) Prenatal exposure to maternal smoking during pregnancy and attention-deficit/hyperactivity disorder in offspring: A meta-analysis. Reprod Toxicol 76:63–70. https://doi.org/10.1016/j.reprotox.2017.12.010 Serati M, Barkin JL, Orsenigo G et al (2017) Research Review: The role of obstetric and neonatal complications in childhood attention deficit and hyperactivity disorder – a systematic review. Child Psychol Psychiatry 58:1290–1300. https://doi.org/10.1111/jcpp.12779 (2017) Attention deficit/hyperactivity disorder and risk of injuries: a systematic review and meta-analysis. J Inj Violence Res 9:. https://doi.org/10.5249/jivr.v9i2.858 Huang A, Wu K, Cai Z et al (2021) Association between postnatal second-hand smoke exposure and ADHD in children: a systematic review and meta-analysis. Environ Sci Pollut Res 28:1370–1380. https://doi.org/10.1007/s11356-020-11269-y Tseng P-T, Yen C-F, Chen Y-W et al (2019) Maternal breastfeeding and attention-deficit/hyperactivity disorder in children: a meta-analysis. Eur Child Adolesc Psychiatry 28:19–30. https://doi.org/10.1007/s00787-018-1182-4 Salazar G, Iniesta R, Bellato A et al (2024) Individualized prediction models in ADHD: a systematic review and meta-regression. Mol Psychiatry. https://doi.org/10.1038/s41380-024-02606-5 Dumitrescu E, Hué S, Hurlin C, Tokpavi S (2021) Machine Learning for Credit Scoring: Improving Logistic Regression with Non-Linear Decision-Tree Effects. Eur J Oper Res. https://doi.org/10.1016/j.ejor.2021.06.053 . 297: Fokkema M, Iliescu D, Greiff S, Ziegler M (2022) Machine Learning and Prediction in Psychological Assessment. Eur J Psychol Assess 38:165–175. https://doi.org/10.1027/1015-5759/a000714 Liu YS, Talarico F, Metes D et al (2024) Early identification of children with Attention-Deficit/Hyperactivity Disorder (ADHD). PLOS Digit Health 3:e0000620–e0000620. https://doi.org/10.1371/journal.pdig.0000620 Gao T, Yang L, Zhou J et al (2024) Development and validation of a nomogram prediction model for ADHD in children based on individual, family, and social factors. J Affect Disord. https://doi.org/10.1016/j.jad.2024.04.069 Garcia-Argibay M, Zhang-James Y, Cortese S et al (2022) Predicting childhood and adolescent attention-deficit/hyperactivity disorder onset: a nationwide deep learning approach. Mol Psychiatry. https://doi.org/10.1038/s41380-022-01918-8 Ter-Minassian L, Viani N, Wickersham A et al (2022) Assessing machine learning for fair prediction of ADHD in school pupils using a retrospective cohort study of linked education and healthcare data. https://doi.org/10.1136/bmjopen-2021-058058 . BMJ Open 12: Nurisa M, Dastghaibyfard G, Hadianfard H (2022) Predicting ADHD from early childhood data using data mining. https://www.researchgate.net/publication/360185624_Predicting_ADHD_from_early_childhood_data_using_data_mining Connelly R, Platt L (2014) Cohort Profile: UK Millennium Cohort Study (MCS). Int J Epidemiol 43:1719–1725. https://doi.org/10.1093/ije/dyu001 Shepherd P, Gilbert E (2019) Millennium Cohort Study Ethical Review and Consent. London: Centre for Longitudinal Studies, UCL Institute of Education. Retrieved from Min X, Li C, Yan Y (2021) Parental Age and the Risk of ADHD in Offspring: A Systematic Review and Meta-Analysis. Int J Environ Res Public Health 18:4939. https://doi.org/10.3390/ijerph18094939 Wilson SV, Cebere B, Myatt J, Wilson S (2022) miceforest (Version 0.6.3). Zenodo. https://doi.org/10.5281/zenodo.742863 Centre for Longitudinal Studies (2020) Millennium Cohort Study User Guide (Surveys 1–5) (9th Edition). London: UCL Institute of Education Liaw A, Wiener M (2002) Classification and Regression by randomForest. Retrieved R News 2 18 Chen T, Guestrin C (2016) XGBoost: a Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD ’16 1:785–794. https://doi.org/10.1145/2939672.2939785 Ke G, Meng Q, Finley T et al (2017) LightGBM: a highly efficient gradient boosting decision tree. Curran Associates Inc., Red Hook, NY, USA, pp 3149–3157 Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20:273–297. https://doi.org/10.1007/BF00994018 Peterson L (2009) K-nearest neighbor. Scholarpedia 4:1883. https://doi.org/10.4249/scholarpedia.1883 Rumelhart DE, Hinton GE, Williams RJ (1986) Learning representations by back-propagating errors. Nature 323:533–536. https://doi.org/10.1038/323533a0 Ketende S, Jones E (2011) The Millenium Cohort Study User Guide to Analysing. MCS Data Using STATA Pedregosa F, Varoquaux G, Gramfort A et al (2011) Scikit-learn: Machine Learning in Python. J Mach Learn Res 12. https://doi.org/10.5555/1953048.2078195 He H, Garcia EA (2009) Learning from Imbalanced Data. IEEE Trans Knowl Data Eng 21:1263–1284. https://doi.org/10.1109/tkde.2008.239 Youden WJ (1950) Index for rating diagnostic tests. Cancer 3:32–35. https://doi.org/10.1002/1097-0142(1950)3:13.0.co;2-3 Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2002) SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res 16:321–357. https://doi.org/10.1613/jair.953 Wollenhaupt-Aguiar B, Librenza-Garcia D, Giovana, Bristot et al (2019) Differential biomarker signatures in unipolar and bipolar depression: A machine learning approach. Psychiatry 54:393–401. https://doi.org/10.1177/0004867419888027 Lundberg S, Lee S-I (2017) A Unified Approach to Interpreting Model Predictions. In: arXiv.org. https://arxiv.org/abs/1705.07874v2 Wang D, Thunéll S, Lindberg U et al (2022) Towards better process management in wastewater treatment plants: Process analytics based on SHAP values for tree-based machine learning methods. J Environ Manage 301:113941. https://doi.org/10.1016/j.jenvman.2021.113941 Spencer NJ, Ludvigsson J, Bai G et al (2022) Social gradients in ADHD by household income and maternal education exposure during early childhood: Findings from birth cohort studies across six countries. PLoS ONE 17:e0264709. https://doi.org/10.1371/journal.pone.0264709 Rabinowicz A, Rosset S (2021) Trees-Based Models for Correlated Data. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2102.08114 Guyon I, Elisseeff A (2003) An introduction to variable and feature selection. J Mach Learn Res. https://doi.org/10.5555/944919.944968 Goodman R (1997) The Strengths and Difficulties Questionnaire: A Research Note. J Child Psychol Psychiatry 38:581–586. https://doi.org/10.1111/j.1469-7610.1997.tb01545.x Iizuka C, Yamashita Y, Nagamitsu S et al (2010) Comparison of the strengths and difficulties questionnaire (SDQ) scores between children with high-functioning autism spectrum disorder (HFASD) and attention-deficit/hyperactivity disorder (AD/HD). Brain Develop 32:609–612. https://doi.org/10.1016/j.braindev.2009.09.009 Croft S, Stride C, Maughan B, Rowe R (2015) Validity of the Strengths and Difficulties Questionnaire in Preschool-Aged Children. Pediatrics 135:e1210–e1219. https://doi.org/10.1542/peds.2014-2920 Sammons P, Sylva S, Melhuish E et al (2004) The Effective Provision of Pre-school Education (EPPE) Project: Technical Paper 11: The continuing effect of pre-school education at age 7 years. Institute of Education, London Melhuish E, Hanna K, Quinn L et al (2004) The Effective Pre-school Provision in Northern Ireland Project, Technical Paper 11: Pre-school Experience and Social/Behavioural Development At the End of Year 3 of Primary School. Stranmillis University, Belfast, N.I. Efron D, Nicholson JM, Anderson V et al (2020) ADHD at Age 7 and Functional Impairments at Age 10. Pediatrics 146:e20201061. https://doi.org/10.1542/peds.2020-1061 Shaw P, Stringaris A, Nigg J, Leibenluft E (2014) Emotion Dysregulation in Attention Deficit Hyperactivity Disorder. Am J Psychiatry 171:276–293. https://doi.org/10.1176/appi.ajp.2013.13070966 Antony EM-A, Pihlajamäki MR, Speyer LG, Murray AL (2021) Does emotion dysregulation mediate the association between ADHD symptoms and internalizing problems? A longitudinal within-person analysis in a large population-representative study. Edinburgh Research Explorer (University of Edinburgh. https://doi.org/10.31234/osf.io/32dky Goodman R (2015) Scoring the Strengths & Difficulties Questionnaire for 2–4 year olds Hartman CA, Luteijn E, Serra M, Minderaa R (2006) Refinement of the Children’s Social Behavior Questionnaire (CSBQ): An Instrument that Describes the Diverse Problems Seen in Milder Forms of PDD. J Autism Dev Disord 36:325–342. https://doi.org/10.1007/s10803-005-0072-z Lundberg SM, Erion G, Chen H et al (2020) From local explanations to global understanding with explainable AI for trees. Nat Mach Intell 2:56–67. https://doi.org/10.1038/s42256-019-0138-9 Claussen AH, Holbrook JR, Hutchins HJ et al (2022) All in the Family? A Systematic Review and Meta-analysis of Parenting and Family Environment as Risk Factors for Attention-Deficit/Hyperactivity Disorder (ADHD) in Children. Prev Sci 25. https://doi.org/10.1007/s11121-022-01358-4 Mohammadian M, Khachatryan LG, Vadiyan FV et al (2025) The association between maternal tobacco smoking during pregnancy and the risk of attention-deficit/hyperactivity disorder (ADHD) in offspring: A systematic review and meta-analysis. PLoS ONE 20:e0317112–e0317112. https://doi.org/10.1371/journal.pone.0317112 Galéra C (2011) Early Risk Factors for Hyperactivity-Impulsivity and Inattention Trajectories From Age 17 Months to 8 Years. Arch Gen Psychiatry 68:1267. https://doi.org/10.1001/archgenpsychiatry.2011.138 Owens J (2021) Parental intervention in school, academic pressure, and childhood diagnoses of ADHD. Soc Sci Med 272:113746. https://doi.org/10.1016/j.socscimed.2021.113746 Sonuga-Barke EJS, Halperin JM (2010) Developmental phenotypes and causal pathways in attention deficit/hyperactivity disorder: potential targets for early intervention? Child Psychol Psychiatry 51:368–389. https://doi.org/10.1111/j.1469-7610.2009.02195.x Yang Y, Shields GS, Zhang Y et al (2022) Child executive function and future externalizing and internalizing problems: A meta-analysis of prospective longitudinal studies. Clin Psychol Rev 97:102194. https://doi.org/10.1016/j.cpr.2022.102194 Marques S, Correia-de-Sá T, Guardiano M et al (2024) Emotion dysregulation and depressive symptoms mediate the association between inhibitory control difficulties and aggressive behaviour in children with ADHD. Front Psychiatry 15. https://doi.org/10.3389/fpsyt.2024.1329401 Lan Y-T, Liu X-P, Fang H-S (2018) Randomized control study of the effects of executive function training on peer difficulties of children with attention-deficit/hyperactivity disorder C subtype. Appl Neuropsychology: Child 9:41–55. https://doi.org/10.1080/21622965.2018.1509003 Willcutt EG, Doyle AE, Nigg JT et al (2005) Validity of the Executive Function Theory of Attention-Deficit/Hyperactivity Disorder: A Meta-Analytic Review. Biol Psychiatry 57:1336–1346. https://doi.org/10.1016/j.biopsych.2005.02.006 Additional Declarations No competing interests reported. Supplementary Files Supplementalmaterials.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7134745","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":489011138,"identity":"da127588-5530-43f2-9181-5c18164a6557","order_by":0,"name":"Zijin Wang","email":"","orcid":"","institution":"Beijing Normal University","correspondingAuthor":false,"prefix":"","firstName":"Zijin","middleName":"","lastName":"Wang","suffix":""},{"id":489011139,"identity":"f4d2dd88-e87c-498e-b003-d6f932b115b7","order_by":1,"name":"Hao Wu","email":"","orcid":"","institution":"Beijing Normal University","correspondingAuthor":false,"prefix":"","firstName":"Hao","middleName":"","lastName":"Wu","suffix":""},{"id":489011140,"identity":"fa7a4c0f-9bdc-4e39-9fb3-ac0833cabd79","order_by":2,"name":"Ruyue Zhai","email":"","orcid":"","institution":"Guizhou Normal University","correspondingAuthor":false,"prefix":"","firstName":"Ruyue","middleName":"","lastName":"Zhai","suffix":""},{"id":489011141,"identity":"a356d948-320c-41dc-aed8-f4e0f2d6c81f","order_by":3,"name":"Libin Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA50lEQVRIiWNgGAWjYDACCRBRIJHAxsB8DCJygCgtBiAtbGkkaWFIYGDgMSNOi/zs5mcPvxhY5PFJ93x7XNjGIMd3I4HxcwEeLYxzjpkbyxhIFLPJnN1uPLONwVjyRgKz9Aw8WpglEsykJQwkEtskcrdJ87YxJG64kcDGzINHC5tE+jeolpxnIC31BLXwSOSYSX6AaGEDaUkwIKRFQiKnTJoB5BeJNHNjnnMShjPPPGyWxqdFfkb6NskfFXV58jOSnz3mKbOR5zuefPAzPi0ggOwMUDQxNhDQAFTyg6CSUTAKRsEoGNEAAG27QIaPS6W4AAAAAElFTkSuQmCC","orcid":"","institution":"Beijing Normal University","correspondingAuthor":true,"prefix":"","firstName":"Libin","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-07-16 01:53:13","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7134745/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7134745/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87701560,"identity":"7cd31d08-1064-46f6-ba0e-7a101705e721","added_by":"auto","created_at":"2025-07-28 07:24:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":135542,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix of six machine learning algorithms. (A) RF Confusion Matrix; (B) SVM Confusion Matrix; (C) LightGBM Confusion Matrix; (D) KNN Confusion Matrix; (F) XGBoost Confusion Matrix; (G) MLP Confusion Matrix.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7134745/v1/400875f5ad6c8bb00d648613.png"},{"id":87702186,"identity":"9ab8cc6e-1e62-4f3b-98e0-9ff4d35a62ff","added_by":"auto","created_at":"2025-07-28 07:32:10","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":250547,"visible":true,"origin":"","legend":"\u003cp\u003eRoc Curves and Precision-Recall Curves of Test set. (A) Roc Curves - Test set; (B) Precision-Recall Curves - Test set.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7134745/v1/3b90f4815ff44da0a3539d7e.png"},{"id":87701552,"identity":"83842868-d835-4e01-911b-f037c54dd009","added_by":"auto","created_at":"2025-07-28 07:24:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":137802,"visible":true,"origin":"","legend":"\u003cp\u003eFeature importance of two machine learning algorithms. (A) XGBoost feature importance; (B) SVM feature importance.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7134745/v1/af16371fd2c40eda77e44a2b.png"},{"id":87701563,"identity":"87fbf9b3-b88c-43c2-a3f3-7bec1cd49832","added_by":"auto","created_at":"2025-07-28 07:24:12","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":68274,"visible":true,"origin":"","legend":"\u003cp\u003eGlobal SHAP summary plot illustrating feature importance in the XGBoost model\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7134745/v1/77693bbfb3634a76274801e2.png"},{"id":87701551,"identity":"d20514b6-fc3e-480d-b0cb-6260eb93aafa","added_by":"auto","created_at":"2025-07-28 07:24:10","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":238415,"visible":true,"origin":"","legend":"\u003cp\u003eLocal feature-value impacts of SHAP analysis for the XGBoost model. (A) Conduct problems; (B) Peer problems; (C) Gender; (D) Emotional-Dysregulation; (E) Breastfeeding; (F) Maternal age at conception 30–39 years; (G) Maternal education level; (H) Maternal age at conception 20–29 years; (I) Single parent; (J) Smoking during pregnancy.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7134745/v1/be3da561176e228872ba7786.png"},{"id":87702194,"identity":"4d6f42db-017c-4718-978c-8152a64acd6a","added_by":"auto","created_at":"2025-07-28 07:32:12","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":32078,"visible":true,"origin":"","legend":"\u003cp\u003eRelative contribution of each domain. CogBehv= Early child cognitive and behavioral development. Demo=Demographic features. Preg=Pregnant features. Parenting=Maternal early parenting style.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7134745/v1/1183caabc5ffddfcd14f7388.png"},{"id":88956317,"identity":"502570a9-6f4b-4f09-9ae2-00e182f9dff5","added_by":"auto","created_at":"2025-08-13 07:10:04","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1501299,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7134745/v1/f98736e1-7536-497e-bf6d-7f5895183f58.pdf"},{"id":87701540,"identity":"f4d0b91d-4393-4a50-81a6-adf2fc5df5fb","added_by":"auto","created_at":"2025-07-28 07:24:10","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":325245,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalmaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-7134745/v1/7fadb7cf2125f13cb5310c26.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Early prediction of childhood ADHD using prenatal and early postnatal behavioural features: evaluation across six machine-learning algorithms","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAttention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder characterised by persistent inattention and/or hyperactivity-impulsivity [1]. Children with ADHD who are identified early and receive timely intervention and treatment have better long-term outcomes [2]. For example, those diagnosed and treated in childhood show higher self-esteem and lower levels of loneliness during adolescence compared to those diagnosed and treated in adolescence [3], as well as a reduced likelihood of delinquent and risky behaviours [4].\u003c/p\u003e\n\u003cp\u003eHowever, early diagnosis is highly challenging in practice. In the UK, national data indicate that although the community prevalence reaches as high as 3\u0026ndash;5% [5], actual clinical recording rates are below 1%, with only 0.51% of children aged 5\u0026ndash;9 years diagnosed [6]. Meanwhile, traditional diagnostic tools have low sensitivity for identifying ADHD in younger children, leading to missed diagnoses [7, 8]. Such underdiagnosis or missed diagnosis of ADHD in early childhood implies that a considerable proportion of high-risk children fail to receive timely intervention. Therefore, relying solely on formal diagnosis before intervention often means missing the optimal window of opportunity for intervention. It is urgently necessary to use risk prediction tools to identify and support high-risk ADHD children early, at symptom onset or before they reach diagnostic thresholds.\u003c/p\u003e\n\u003cp\u003eOn the other hand, a recent systematic meta-review of systematic reviews on ADHD demonstrated that ADHD risk factors span both the prenatal period and the early postnatal years (here broadly defined as the period after birth) and involve biological, behavioural and environmental domains [9]. In the prenatal phase, such as maternal overweight [10], smoking during pregnancy [11], and low birth weight [12] have all been linked to a higher risk of offspring ADHD; in the postnatal phase, for example, injuries [13] and second-hand smoke [14] elevate ADHD risk, whereas breastfeeding appears protective [15]. These findings suggest that simultaneously incorporating prenatal and postnatal multidomain features\u0026mdash;and thereby integrating information across developmental periods\u0026mdash;may potentially improve early prediction of ADHD in children.\u003c/p\u003e\n\u003cp\u003eIn recent years, many multivariable, multi-domain machine-learning (ML) models for ADHD prediction have proliferated [16]. Compared with traditional approaches (e.g., logistic regression), ML models not only can predict ADHD just as effectively before clinical-threshold symptom onset but also are better suited to integrating heterogeneous, multidomain data. Especially non-linear ML algorithms automatically capture higher-order and non-linear feature interactions, mitigating multicollinearity and thereby offering higher accuracy and greater practical feasibility for early ADHD prediction [17, 18].\u003c/p\u003e\n\u003cp\u003eNonetheless, several limitations persist in current early-prediction studies for childhood ADHD. First, to our knowledge, only a handful of ML models have simultaneously considered risk factors from both prenatal and postnatal domains [19\u0026ndash;21]. Even these few studies include only limited antenatal variables and do not evaluate the relative contribution of prenatal versus postnatal exposures to model performance. Moreover, among these studies, only one investigation has attempted to predict ADHD truly early in childhood, leveraging features measured before age 5 to forecast ADHD at 5\u0026ndash;6 years [19]. In contrast, the remaining studies examine broader age ranges from childhood to adolescence.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMoreover, UK-based ML studies that specifically target paediatric ADHD are exceedingly scarce, with existing studies primarily focusing on predicting ADHD at ages 6\u0026ndash;7 based on risk factors identified at ages 4\u0026ndash;5 [22]. To date, no research has explored even earlier prediction within UK cohorts (e.g. using factors measured at or before age three). Although a recent preprint attempted to develop an early ADHD prediction model for British children [23], approximately 98% of healthy samples in the initial dataset were removed to balance ADHD and healthy samples, potentially inducing a prior-probability shift and somewhat inflating model performance in real-world cohorts.\u003c/p\u003e\n\u003cp\u003eGiven the limitations of previous studies described above, this study based on the UK Millennium Cohort Study (MCS) to predict ADHD outcomes at early school age (5\u0026ndash;7 years) using information collected before age three. Four objectives are pursued: (1) after combining early postnatal features (0\u0026ndash;3 years) with a broad set of maternal prenatal features, we will compare six machine-learning algorithms (Random Forest, Support Vector Machine, Light Gradient Boosting Machine, K-Nearest Neighbors, Extreme Gradient Boosting, and Multilayer Perceptron) and select the optimal model, thereby addressing the gap in UK research on predicting ADHD using information available prior to age three; (2) enhance model generalisability by applying sample weights for ethnic-minority and low-income groups; (3) employ SHAP analysis to provide an interpretable account of the key predictors identified by the optimal model; and (4) use SHAP values to assess the relative contributions of different domains to overall model performance.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eParticipants and design\u003c/p\u003e\n\u003cp\u003eThe dataset used in this study is the Millennium Cohort Study (MCS). The MCS aims to track the development of UK children born in 2000 or 2001 [24]. Data collection methods include questionnaires and interviews, which continuously gather information on the health, cognitive development, educational achievement, and socioeconomic background of cohort members. However, the sample structure of the MCS has limitations, including geographically uneven distribution and disproportionate representation of ethnic minorities and families from disadvantaged areas. Therefore, clustering, stratification, and weighting methods are required to post hoc correct model parameters and balance population representativeness [24]. All cohort members participating in this study provided informed consent signed by themselves or by their caregivers acting as proxies [25].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDefinition of outcome and Sample derivation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe present study selected early ADHD diagnosis data as the outcome, at age 5 (sweep 3) and age 7 (sweep 4). To better define the composition of the early ADHD diagnostic group, participants who were not diagnosed in early childhood but received an ADHD diagnosis at age 11 (sweep 5) or age 14 (sweep 6) (i.e., late childhood and early adolescence) were excluded, thus controlling for their potential confounding effect as early \u0026ldquo;healthy controls\u0026rdquo; in the healthy group. The early and late ADHD diagnostic data were obtained from four datasets within the MCS: mcs3_parent_cm_interview and mcs4_parent_cm_interview, as well as mcs5_parent_cm_interview and mcs6_parent_cm_interview, respectively. The variables CPADHD00 and DPADHD00 (reported by mothers) constituted the early ADHD diagnosis, whereas the variables EPADHD00 and FPADHD00 constituted the late ADHD diagnosis (i.e., whether diagnosed with ADHD).\u003c/p\u003e\n\u003cp\u003eGiven missing data at these four time points, the following rules were applied to select participants for early/late ADHD diagnostic groups:\u003c/p\u003e\n\u003cp\u003e1) For early ADHD diagnosis, participants who reported a confirmed ADHD diagnosis at either age 5 or age 7 (i.e., CPADHD00 or DPADHD00 = 1) were classified into the ADHD diagnostic group. Participants who reported no ADHD diagnosis at both ages 5 and 7 (i.e., CPADHD00 and DPADHD00 = 2) were classified into the healthy group. Participants who did not meet either criterion were marked as missing.\u003c/p\u003e\n\u003cp\u003e2) For late ADHD diagnosis, participants who reported a confirmed ADHD diagnosis at either age 11 or age 14 (i.e., EPADHD00 or FPADHD00 = 1) were classified into the ADHD diagnostic group. Participants who reported no ADHD diagnosis at both ages 11 and 14 (i.e., EPADHD00 and FPADHD00 = 2) were classified into the healthy group. Participants who did not meet either criterion were marked as missing.\u003c/p\u003e\n\u003cp\u003eThe samples for the early/late diagnostic groups identified based on the above rules underwent further selection procedures, as detailed in Table 1. Sweep 3 initially included 15,373 cohort members, of whom 9,385 remained after screening. Among these, 9,126 participants were classified into the healthy control group without an ADHD diagnosis, and 259 (2.76%) were classified into the ADHD diagnostic group.\u003c/p\u003e\n\u003cp\u003eTable 1.\u0026nbsp;Sample selection pipeline.\u003c/p\u003e\n\u003cp\u003e\u003cimg width=\"478\" height=\"534\" src=\"data:image/png;base64,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\" alt=\"图示AI 生成的内容可能不正确。\"\u003e\u003c/p\u003e\n\u003cp\u003eData collection\u003c/p\u003e\n\u003cp\u003eThe time range of feature selection in this study spans from the mother\u0026apos;s early pregnancy to the child\u0026apos;s third birthday. Apart from demographic features, the features in this study were divided into two main domains based on the child\u0026rsquo;s birth: prenatal features and early postnatal features (i.e. early childhood period, ages 0-3). Features with more than 10% missing values were excluded. Additionally, paternal ADHD-related features were entirely excluded due to substantial missing data in the MCS; therefore, this study only included mother-reported features.\u003c/p\u003e\n\u003cp\u003eUltimately, a total of 45 features were included in this study. Four demographic features were included: Gender, Race, Household income quintiles, and Mothers academic qualification. Fifteen prenatal features were included: Premature, Low weight, Prenatal depression, Prenatal eclampsia, Prenatal epilepsy, Prenatal Suspected Slow Growth, Multiple pregnancy, Threatened miscarriage, Prenatal asthma, Prenatal persistent vomiting, Prenatal non-trivial infections, Prenatal smoking, Prenatal drinking, Age group, and Prepregnancy BMI. Twenty-six early postnatal features were included and further divided into three sub- domains: (1) Early maternal parenting style: Mother-child Conflicts, Mother-child positive relationship, Breastfeeding, Postpartum depression, Regular bedtimes, Regular eat, Single parent; (2) Child environmental exposure: Media, Injury, Secondhand smoke; (3) Early child cognitive and behavioural development: Suspected social delay, Suspected fine motor delay, Suspected gross motor delay, Mood, Approach, Adaptability, Regularity, Cry, Naming vocabulary, School readiness composit, Independence-Self Regulation, Emotional-Dysregulation, Emotional Symptoms, Conduct problems, Peer problems, Prosocial Behaviour. For more details, see Table A1in the Appendix.\u003c/p\u003e\n\u003cp\u003eGiven that maternal age during pregnancy showed a nonlinear relationship with offspring ADHD [26], Age group was one-hot encoded similarly to Race, using \u0026quot;Pregnant at the age of 12-19\u0026quot; as the baseline level. After one-hot encoding, the total number of features increased to 53.\u003c/p\u003e\n\u003cp\u003eData preprocessing\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor features with less than 10% missing data, nonlinear multiple imputation was performed using miceforest 0.6.3 [27] in Python 3.12.3. Missing values for the Sweep 4 weighting variable (i.e., DOVWT2) were set to 1, indicating that they did not participate in weight adjustment [28]. The Dataset was randomly divided using train_test_split, with a fixed random seed of 42 and a 7:3 ratio between the training and test sets.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eStatistical analysis\u003c/p\u003e\n\u003cp\u003eThis study evaluated six nonlinear ML algorithms, including tree-based models Random Forest (RF) [29], Extreme Gradient Boosting (XGBoost) [30], Light Gradient Boosting Machine (LightGBM) [31]; and non-tree-based models Support Vector Machine (SVM) [32], K-Nearest Neighbors (KNN) [33], Multilayer Perceptron (MLP) [34]. The rationale for selecting these algorithms was that the dataset had a relatively large sample size and contained highly nonlinear features, including binary, categorical, and continuous variables. RF, XGBoost, KNN, LightGBM, and MLP can directly handle high-domain nonlinear data, while the SVM utilises a Radial Basis Function (RBF) kernel to handle such nonlinear relationships. ML analyses were implemented using Python 3.7 and Scikit-learn 0.24.1. The code is freely available at GitHub: https://github.com/Anankysus/Early-ADHD-ML/tree/main.\u003c/p\u003e\n\u003cp\u003eFollowing the User Guide to Analysing MCS Data using Stata [35], the primary sampling unit variable SPTN00 and the overall sample weight DOVWT2 (national sampling weights based on Sweep 4) were selected for model development and evaluation, thus accounting for clustering correlations and sampling probability differences.\u003c/p\u003e\n\u003cp\u003eIn this study, the ratio of ADHD diagnostic cases to healthy controls was 1:35.3, reflecting extreme class imbalance due to the predominance of the healthy group. Therefore, an additional validation set was not created; instead, cross-validation was used. To mitigate the issue of ML models excessively focusing on the majority (healthy) class due to this imbalance, three balancing strategies were employed: Class Weights [36], Random Undersampler [37], and threshold tuning using Youden\u0026rsquo;s J index [38]. Specifically, the Random Undersampler ratio was set to 0.2 (random seed = 42), resulting in a healthy-to-ADHD ratio of 905:181 in the training set. Empirical testing showed that a ratio of 0.2 not only yielded optimal model performance but also outperformed oversampling methods such as SMOTE and SMOTE-ENN [39].\u003c/p\u003e\n\u003cp\u003eTo minimise the potential impact of redundant features on ML models, Recursive Feature Elimination (RFE) [40] was employed for feature selection in each of the six models to determine the optimal subset of features. RFE and model training were embedded in the same pipeline, where both feature selection and hyperparameter tuning (GridSearchCV) were conducted simultaneously within the same cross-validation framework. All ML hyperparameter settings are detailed in Table A2 in the Appendix. Within this pipeline, to further avoid information leakage among samples within the primary sampling unit (PSU) during hyperparameter tuning, Group K-Fold cross-validation (with MCS\u0026rsquo;s cluster identifier SPTN00 as the grouping factor) was conducted on the training set. Specifically, a 10-fold group-wise cross-validation was used, with 9-folds for training and 1-fold for validation in each iteration. The random seed was fixed at 42 for all models. After completing the training phase, each base classifier\u0026rsquo;s raw prediction scores were calibrated using 10-fold sigmoid calibration to correct model confidence biases caused by extreme class imbalance.\u003c/p\u003e\n\u003cp\u003eTo ensure unbiased estimation of all performance metrics according to the target population, complex survey sampling weights from MCS were incorporated, along with class weights, to form composite weights integrated into each ML model. Performance metrics were calculated using three averaging approaches: macro-averaging, micro-averaging, and weighted averaging. Macro-averaging calculates metrics independently per class and then averages these; micro-averaging merges predictions and true labels across classes before calculating overall metrics; weighted averaging computes metrics per class and then averages these metrics according to each class\u0026rsquo;s composite weight in the test set. Detailed evaluation metrics included Accuracy, Precision, Specificity, Recall, and F1-score.\u003c/p\u003e\n\u003cp\u003eReceiver Operating Characteristic (ROC) curves and corresponding Area Under the Curve (AUC) were used to measure the overall model discrimination between positive and negative classes across all thresholds. Additionally, Precision\u0026ndash;Recall (PR) curves were plotted, which emphasise the trade-off between recall and precision for the minority class (ADHD group) in highly imbalanced datasets, complementing the ROC curves that may lack sensitivity for minority-class performance. Confusion matrices were generated for each ML model to illustrate the distribution of true positives, false positives, true negatives, and false negatives, aiding identification of specific misclassification patterns. Composite weights were also incorporated into ROC, confusion matrix, and PR curve analyses to mitigate biases resulting from imbalanced classes and sampling design.\u003c/p\u003e\n\u003cp\u003eFeature importance was reported for the model with the best test performance, reflecting each variable\u0026rsquo;s overall contribution to model performance (taking composite weights into account). However, feature importance only indicates the critical features influencing model discrimination, without clarifying whether their effects are positive or negative at various values. Therefore, SHAPley Additive exPlanations (SHAP) [41, 42] were employed for detailed analysis of the best-performing model. SHAP decomposes the model\u0026rsquo;s output into additive feature contributions, enabling both a global feature ranking (by cumulative contributions) and local analysis of each feature\u0026rsquo;s specific values, enhancing model transparency and reliability [41]. The global SHAP feature influences were presented as violin plots, using replicated samples to simulate composite weights, while local feature-value impacts based on SHAP values were illustrated using boxplots Lastly, cross- domain group comparisons of contributions to the model were performed based on SHAP values.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eClinical characteristics\u003c/p\u003e\n\u003cp\u003eDemographic features of this study are shown in Table 2. Among the total sample of 9385 cohort members, there were 4769 girls (50.8%) and 4616 boys (49.2%). Ethnicity was predominantly White (85.0%). The distribution of participants across Household income quintiles was relatively even, with the Lowest quintile having the smallest proportion (15.17%) and the other four quintiles ranging between 20% and 21%. Regarding Mothers academic qualification, the largest group was O level/GCSE grades A-C (30.5%), while the smallest proportion was Higher degree (3.8%).\u003c/p\u003e\n\u003cp\u003eWithin specific groups, gender was evenly distributed in the healthy group but severely imbalanced in the ADHD diagnostic group, with a female-to-male ratio of approximately 1:4. White ethnicity dominated both healthy and ADHD groups, slightly more so in the ADHD diagnostic group. The distribution of Household income weighted quintiles was relatively balanced in the healthy group; however, the lowest-income group dominated (39.0%) in the ADHD diagnostic group. Additionally, mothers with higher academic qualifications were generally less represented in the ADHD diagnostic group compared to the healthy group.\u003c/p\u003e\n\u003cp\u003eTable 2. Participant demographic features.\u003c/p\u003e\n\u003ctable width=\"557\"\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003eSample characteristic\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003eNo. (%)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"44\"\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"197\"\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"116\"\u003e\n\u003cp\u003eADHD diagnosed\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003eHealthy\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\n\u003cp\u003eTotal\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"44\"\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003eNo. of samples\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"70\"\u003e\n\u003cp\u003e259\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\n\u003cp\u003e9126\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\n\u003cp\u003e9385\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"44\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"197\"\u003e\n\u003cp\u003eGender\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"70\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"44\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"197\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Girl\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e50(19.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e4719(51.7)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e4769(50.8)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"197\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Boy\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e209(80.7)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e4407(48.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e4616(49.2)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Missing\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"70\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"44\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"197\"\u003e\n\u003cp\u003eRace\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"70\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"44\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; White\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e231(89.2)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e7764(84.9)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e7977(85.0)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Mixed\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e8(3.1)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e243(2.7)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e251(2.7)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Indian\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e5(1.9)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e227(2.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e232(2.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Pakistani and Bangladeshi\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e8(3.1)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e545(6.0)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e553(5.9)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Black or Black British\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e5(1.9)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e247(2.7)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e252(2.7)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Other Ethnic group\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e2(0.8)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e115(1.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e117(2.7)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Missing\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"70\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e3(0.0003)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e3(0.0003)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003eHousehold income quintiles\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"70\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"44\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Lowest quintile\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e101(39.0)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e1568(17.2)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e1669(17.8)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Second quintile\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e69(26.6)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e1839(20.2)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e1908(20.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Third quintile\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e36(13.9)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e1865(20.4)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e1901(20.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Fourth quintile\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e30(11.6)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e1921(21.0)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e1951(20.8)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Highest quintile\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e23(8.9)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e1923(21.1)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e1946(20.7)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Missing\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"70\"\u003e\n\u003cp\u003e0\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e10(0.1)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e10(0.1)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003eMothers academic qualification\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"70\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"39\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd width=\"44\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Higher degree\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e5(1.9)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e350(3.8)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e355(3.8)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; First degree\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e24(9.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e1317(14.4)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e1341(14.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Diplomas in higher education\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e19(7.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e849(9.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e868(9.2)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; A/AS/S levels\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e24(9.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e867(9.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e891(9.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; O level/GCSE grades A-C\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e75(29.0)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e2785(30.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e2860(30.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; GCSE grades D-G\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e17(6.6)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e842(9.2)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e859(9.2)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Other academic qualifications\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e6(2.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e233(2.6)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e239(2.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"243\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; None of these qualifications\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e52(20.1)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e1303(14.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e1355(14.4)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd width=\"220\"\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp; Missing\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd width=\"77\"\u003e\n\u003cp\u003e37(14.3)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"77\"\u003e\n\u003cp\u003e580(6.4)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" width=\"83\"\u003e\n\u003cp\u003e617(6.5)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003ePredictive performance\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe optimal hyperparameters for the six ML models are presented in Table A3 in the Appendix. After applying RFE, each model retained 22 variables. Table 3 presents the weighted accuracy scores for the training set and testing set of each ML model. On the test set, XGBoost achieved the highest weighted accuracy (0.8126), while the weighted accuracies of the other five models ranged from 0.7 to 0.8, with LightGBM having the lowest weighted accuracy (0.7087). Comparing weighted accuracy between the training and testing sets, RF and LightGBM exhibited potential overfitting, with a gap greater than 10%.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u003c/strong\u003e\u0026nbsp; Weighted accuracy Scores of Training Set and Testing Set for six machine learning algorithms\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"576\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003eAlgorithms\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"bottom\" style=\"width: 364px;\"\u003e\n \u003cp\u003eADHD diagnosed group vs Healthy group\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 229px;\"\u003e\n \u003cp\u003eTraining set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 135px;\"\u003e\n \u003cp\u003eTest set\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 53px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 229px;\"\u003e\n \u003cp\u003e0.9346\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 135px;\"\u003e\n \u003cp\u003e0.7918\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 53px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 229px;\"\u003e\n \u003cp\u003e0.7882\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 135px;\"\u003e\n \u003cp\u003e0.7954\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003eLightGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 53px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 229px;\"\u003e\n \u003cp\u003e0.9659\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 135px;\"\u003e\n \u003cp\u003e0.7087\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 53px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 229px;\"\u003e\n \u003cp\u003e0.7716\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 135px;\"\u003e\n \u003cp\u003e0.7763\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 53px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 229px;\"\u003e\n \u003cp\u003e0.8453\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 135px;\"\u003e\n \u003cp\u003e0.8126\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003eMLP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 53px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 229px;\"\u003e\n \u003cp\u003e0.7845\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 135px;\"\u003e\n \u003cp\u003e0.7945\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 4 presents the evaluation metrics for the six models based on macro-, micro-, and weighted-averaging approaches, and Fig. 1 displays the weighted confusion matrix results for each model. After accounting for sample frequency and sampling weights, weighted F1 was adopted as the primary evaluation metric, supplemented by weighted precision, weighted recall, macro F1, and diagnostic recall rate. XGBoost achieved a weighted F1 score of 0.813, a weighted precision of 0.820, a weighted recall of 0.813, and a macro F1 score of 0.585 on the test set. Its confusion matrix showed a diagnostic recall rate of 76.3%. Considering these comprehensive metrics, XGBoost outperformed all other models. Combined with the highest weighted accuracy, XGBoost was selected as the preferred model due to its superior classification performance and robustness to extreme class imbalance. Its optimal threshold was 0.1300, with a Youden\u0026rsquo;s J index of 0.611.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4\u003c/strong\u003e. Six machine learning algorithms performance for ADHD diagnosed\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"600\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\n \u003cp\u003eAlgorithms\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003eF1 score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emacro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9270\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5963\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7904\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.6336\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emicro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9270\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9194\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9194\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9194\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003eweighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9203\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8180\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7918\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7901\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emacro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8327\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5540\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8010\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5534\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emicro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8327\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8310\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8310\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8310\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003eweighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7975\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7954\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7956\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\n \u003cp\u003eLightGBM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emacro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9383\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5884\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7255\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.6200\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emicro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9383\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9265\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9265\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9265\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003eweighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.9367\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7740\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7087\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.6977\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emacro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7779\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5398\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7672\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5150\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emicro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7779\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7773\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7773\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7773\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003eweighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7651\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7766\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7763\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7764\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emacro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8795\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5667\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7923\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5846\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emicro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8795\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8746\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8746\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8746\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003eweighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8704\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8201\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8125\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8126\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\n \u003cp\u003eMLP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emacro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7893\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5458\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7985\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.5275\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003emicro\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7893\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7898\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7898\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7898\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 84px;\"\u003e\n \u003cp\u003eweighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7708\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7943\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7945\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7944\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAside from LightGBM, the weighted accuracy of the remaining four models did not differ significantly from that of XGBoost (Holm-adjusted p-value \u0026gt; 0.05) (See Table A4 in the Appendix). However, confusion matrices indicated limitations and shortcomings in suppressing false positives and false negatives in each model. Among these, SVM had a minority-class recall rate of 78.0%, demonstrating slightly higher sensitivity and thus a lower risk of false negatives compared to XGBoost, but at the cost of approximately a 6% increase in false-positive rate. Nevertheless, due to its recall rate of over 80% for true positives, SVM was identified as the second-best model.\u003c/p\u003e\n\u003cp\u003eROC and PR curves further validated the robust predictive performance of XGBoost in the classification task (see Fig. 2). XGBoost achieved the highest AUC (0.881) and the second-highest Average Precision (AP = 0.900) on the test set, with both curves maintaining high specificity and precision at high recall levels. In contrast, although RF achieved the highest AP (0.908) and second-highest AUC (0.879) on the PR curve, these values were inflated by its exceptionally high recall of the healthy group (92.03%), whereas its diagnostic recall rate was the second lowest (68.11%) among the six models (see Fig. 1A). Although SVM, LightGBM, KNN, and MLP were not selected as preferred models, they all obtained AUC values greater than 0.84 and AP values exceeding 0.86, demonstrating similarly high accuracy for overall classification.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFeature importance\u003c/p\u003e\n\u003cp\u003eFig. 3A shows the ranking of feature importance for the optimal XGBoost model. Early child cognitive and behavioural development features dominated numerically, comprising 10 features (approximately 45.5%): Conduct problems, Peer problems, Prosocial behaviour, School readiness assessment, Emotional-Dysregulation, Naming vocabulary, Cry, Regularity, Adaptability, and Approach. There were 5 prenatal features: Age_group_2 (Pregnant at the age of 20-29), Prenatal smoking, Age_group_3 (Pregnant at the age of 30-39), Prepregnancy BMI, and Prenatal drinking; Four early maternal parenting style features: Breastfeeding, Single parent, Regular bedtimes, and Regular eat; Two demographic features: Gender and Mothers academic qualification; and 1 children environmental exposure feature: Secondhand smoke.\u003c/p\u003e\n\u003cp\u003eGiven that tree-based and non-tree-based models differed in terms of feature selection and ranking of feature importance, and given that the performance differences between models were not significant in this study (See Fig. A1 in the Appendix), the present study additionally examined the feature composition of non-tree-based models, using SVM as the best-performing non-tree-based model (see Fig. 3B). Among the 15 features positively contributing to the SVM model, only 3 belonged to early child cognitive and behavioural development: Conduct problems, Emotional-Dysregulation, and Suspected social delay. In contrast, prenatal features accounted for the largest proportion (approximately 45.3%), including Threatened miscarriage, Prenatal non-trivial infections, Premature, Prenatal smoking, Low weight, and Age_group_3 (Pregnant at the age of 30-39). Additionally, there were 3 demographic features: Race_4 (Pakistani and Bangladeshi), Gender, and Race_5 (Black or Black British); 1 environmental exposure feature: Secondhand smoke; and 2 early maternal parenting style features: Breastfeeding and Postpartum depression.\u003c/p\u003e\n\u003cp\u003eOverall, prenatal factors accounted for approximately 22.7% of the selected features in the XGBoost model. In contrast, prenatal factors accounted for approximately 45.2% of the effective features in SVM, the representative non-tree-based model. Thus, the proportion of prenatal features retained in SVM exceeded that of XGBoost.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo further explore the feature selection preferences of these two models regarding prenatal and postnatal domains, two sensitivity analyses were conducted: a linear correlation analysis between prenatal and early postnatal feature groups and a permutation-based sensitivity analysis using 10-fold cross-validation (see Table A5 \u0026amp; Table A6 in the Appendix). The linear correlation analysis revealed minimal linear redundancy and high independence between these two feature groups, with a mean absolute correlation coefficient (|\u0026rho;|) of 0.0301 and a 75th percentile of less than 0.037, indicating that very few pairs exceeded a correlation of 0.10. The permutation sensitivity analysis showed that for XGBoost, the average AUC decreased by only 0.0129 after permutation of prenatal features, whereas it decreased by 0.2440 after permutation of postnatal features. In contrast, for SVM, average AUC decreased by 0.0274 after permutation of prenatal features\u0026mdash;approximately 2.1 times greater than for XGBoost\u0026mdash;and decreased by 0.1756 after permutation of postnatal features, approximately 0.72 times that of XGBoost.\u003c/p\u003e\n\u003cp\u003eSHAP Analysis of XGBoost\u003c/p\u003e\n\u003cp\u003eThis study presents the SHAP value distributions of the 10 features with the highest contributions to the model output (see Fig. 4), collectively explaining 91% of the model\u0026rsquo;s predictive output, thus dominating overall prediction. Features are sorted from top to bottom based on mean absolute SHAP values. The horizontal axis represents the SHAP value, where positive SHAP values indicate a feature pushing predictions toward the ADHD diagnostic group (positive class), and negative SHAP values indicate predictions toward the healthy group (negative class). Each violin plot\u0026rsquo;s colour represents the original feature values, with red indicating higher feature values and blue indicating lower feature values.\u003c/p\u003e\n\u003cp\u003eIn general, for continuous features (Conduct problems, Peer problems, Emotional-Dysregulation), higher feature values were associated with increased SHAP values, driving predictions toward the ADHD diagnostic group. Among binary features, lower Gender values (i.e., male) predicted the diagnostic group, whereas higher Age_group_3 values (Yes) predicted the healthy group. High values for other binary features such as Breastfeeding (No), Age_group_2 (Yes), Single parent (Yes), and Prenatal smoking (Yes) also predicted classification in the ADHD diagnostic group. Finally, the ordinal feature Mothers academic qualification showed that higher academic qualification predicted classification in the diagnostic group.\u003c/p\u003e\n\u003cp\u003eBased on the global SHAP summary plot, this study further examined local feature-value impacts, illustrating how specific values within each of the top 10 features influenced diagnostic predictions (see Fig.5). Specifically, the feature trends for binary variables (i.e., Gender, Age_group_3, Breastfeeding, Age_group_2, Single parent, and Prenatal smoking) aligned with the global SHAP results. For continuous variables, higher scores on Conduct problems (\u0026gt;3 points, range 0\u0026ndash;10), Peer problems (\u0026gt;2 points, range 0\u0026ndash;10), and Emotional-Dysregulation (\u0026gt;1.57 points, range 1\u0026ndash;3) were more predictive of ADHD diagnosis.\u003c/p\u003e\n\u003cp\u003eRegarding the ordinal variable, Mothers academic qualification, qualifications higher than O level/GCSE grades A\u0026ndash;C were more predictive of ADHD diagnosis. This result contradicts findings from multinational cohorts and previous studies based on the UK MCS dataset [43]. To confirm that this reversal was not a statistical artefact, two sensitivity analyses were conducted (see Table A7 \u0026amp; Fig. A2 in the Appendix). Univariate logistic regression indicated that each decrease in educational qualification level corresponded to an OR = 1.14 (p \u0026lt; 0.001), consistent with prior studies. Additionally, removing sampling weights from the SHAP analysis did not alter the direction of this feature, indicating that its association was not influenced by sampling weights.\u003c/p\u003e\n\u003cp\u003eFinally, using cumulative absolute SHAP values, the contributions of the ten most important features across four predefined domains were quantified and compared, together these domains account for 100 % of the total contribution (Fig. 6). The domain of early child cognitive and behavioural development contributed over half (51.9%), whereas domain of demographic features, prenatal features, and maternal early parenting style contributed comparably, accounting for 19.8%, 14.7%, and 13.6% respectively.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study predicts early childhood ADHD at ages 5\u0026ndash;7 based on features collected before age 3, combining prenatal and postnatal multidomain features. Among the six evaluated models, XGBoost exhibited the best overall performance, achieving an AUC of 0.881. This places the model near excellent performance among ADHD diagnostic models (where reported AUCs range between 0.50 and 0.96) and ranks highly even among prediction models relying solely on behavioural data [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Moreover, the confusion matrix for XGBoost indicates that the model effectively identifies ADHD cases while minimising false negatives.\u003c/p\u003e\u003cp\u003eRegarding feature composition, considering that the five-item Hyperactivity/Inattention (H/I) subscale from the SDQ is similar with DSM-IV/V ADHD diagnostic criteria, including this subscale in the model would likely result in symptom-overlap effects, inflating internal test performance and reducing the model\u0026rsquo;s capacity to detect other early risk signals. To emphasize genuine predictive value rather than simply \"earlier administration of a diagnostic scale,\" this study excluded the H/I subscale, despite the potential slight decrease in sensitivity. Nevertheless, all models in this study maintained good-to-excellent predictive capability.\u003c/p\u003e\u003cp\u003eThe feature importance results revealed differences between XGBoost and SVM in their selection of prenatal and postnatal features, suggesting that SVM may show a stronger preference for prenatal factors. Further sensitivity analyses demonstrated that, even under conditions of weak correlations between prenatal and postnatal feature groups, XGBoost tended to prioritize only a limited number of high-gain postnatal features [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. In contrast, SVM\u0026rsquo;s RBF kernel distance mechanism enabled it to incorporate weaker but stable signals from both prenatal and postnatal features [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. These results further support the idea that SVM favours prenatal features, whereas XGBoost focuses predominantly on a smaller subset of postnatal features, providing the highest gain. Such findings have implications for future development of early ADHD prediction tools: the choice of algorithm itself may influence the preference for predictive features across different time frames and multidomain domains. Further research is needed to validate these algorithm-specific preferences in other datasets.\u003c/p\u003e\u003cp\u003e\u003cb\u003eSHAP Analysis of XGBoost\u003c/b\u003e\u003c/p\u003e\u003cp\u003eA major strength of this study is the SHAP analysis performed on the XGBoost model, which provided a comprehensive evaluation and interpretation of how specific feature values affect ADHD prediction. Conduct problems from the Strengths and Difficulties Questionnaire (SDQ) [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e] contributed the highest mean marginal impact on the model output. Notably, Conduct problems ranked distinctly first in terms of feature importance, a finding consistently observed across other models as well (see Fig. A1 in the Appendix). Based on these findings, this study suggests that early childhood Conduct problems may serve as one of the most critical behavioural indicators for predicting subsequent ADHD. Existing literature supports this hypothesis: previous SDQ-based studies have found that children with ADHD consistently exhibit elevated Conduct problems [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e], and multivariate analysis of SDQ subscales at age 3 further demonstrated that, apart from the H/I subscale, Conduct problems is the only item independently predictive of ADHD diagnosis two years later [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e].\u003c/p\u003e\u003cp\u003ePeer problems from the SDQ [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e] and Emotional-Dysregulation from the Child Social Behaviour Questionnaire (CSBQ) [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e] also significantly contributed to predicting subsequent ADHD, though their impacts were not as pronounced as Conduct problems. Previous longitudinal cohort studies have reported that children with ADHD score significantly higher on Peer problems compared to healthy populations, a trend that persists longitudinally [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. Emotional-Dysregulation is similarly prevalent among children with ADHD, with meta-analyses indicating it may affect approximately 25\u0026ndash;45% of this population [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e]. Interestingly, although Emotional Symptoms from the SDQ [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e] was also considered as a candidate emotional feature, it proved consistently less crucial than Emotional-Dysregulation across all models, both in feature importance rankings and SHAP analyses. This suggests Emotional-Dysregulation is more closely aligned with core ADHD processes. Prior findings also support this interpretation, indicating that Emotional-Dysregulation acts as a mediator between ADHD and Emotional Symptoms. After controlling for this mediating effect, the direct influence of ADHD on Emotional Symptoms is not significant [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eFurther, local feature-value impacts based on SHAP values indicated that SHAP contributions for Conduct problems, Peer problems, and Emotional-Dysregulation tended to plateau at lower scores, stabilising marginal SHAP values. For Conduct problems and Peer problems, the inflexion points in SHAP values were below the \"Borderline\" thresholds (approximately top 20% in the normative data) established by SDQ for UK children aged 2\u0026ndash;4 years [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e], suggesting that the model was able to identify symptom levels below official clinical cut-offs as potential early-risk signals. For Emotional-Dysregulation, despite the absence of a formal cut-off [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e], the SHAP inflexion point (1.862) was below the sample mean in this study, with approximately 75% of children scoring above this inflection point. In fact, these apparently \"low-score, high-risk\" cases are not misclassifications. XGBoost models do not rely on predefined thresholds or percentile ranks; instead, they identify feature splits with the maximum information gain directly from the training data [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Thus, samples below established scale thresholds or percentiles may trigger secondary splits in decision trees, generating positive SHAP contributions and predicting high-risk status [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e]. Furthermore, approximately 21% and 29% of children in this study scored\u0026thinsp;\u0026ge;\u0026thinsp;3 for Conduct problems and \u0026ge;\u0026thinsp;2 for Peer problems, respectively. Although these scores do not strictly surpass the borderline threshold (20%), they closely align with the theoretical proportion for the \"Borderline\" range on the SDQ, indicating these children represent a borderline risk zone warranting greater practical attention.\u003c/p\u003e\u003cp\u003eSHAP analysis results for several binary variables were consistent with previous studies, indicating that being male and not receiving breastfeeding predicted children's subsequent risk of ADHD, a conclusion supported by a recent systematic meta-review of systematic reviews on ADHD [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]; similarly, single parent and prenatal smoking have similarly been shown by previous systematic reviews and meta-analyses to predict ADHD [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e, \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eCompared with the baseline category \"Pregnant at the age of 12\u0026ndash;19,\" pregnancy at the age of 30\u0026ndash;39 was more predictive of the healthy group, consistent with previous meta-analysis findings [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. In contrast, pregnancy at the age of 20\u0026ndash;29 tended to predict the ADHD group, contradicting the commonly reported conclusion in the literature that pregnancy before age 20 carries the highest risk. This discrepancy may result from RFE excluding the \"Pregnant at the age of 40\u0026ndash;49\" category, causing both the 12\u0026ndash;19 and 40\u0026ndash;49 age categories to be coded as zero in the dummy-variable matrix, thus forming a composite baseline. Although RFE improved model performance, it reduced the interpretability of SHAP values for the maternal age variable. Therefore, SHAP results for pregnancy at the age of 20\u0026ndash;29 should be interpreted cautiously.\u003c/p\u003e\u003cp\u003eRegarding Mothers academic qualification, since sampling design factors were excluded, it can be inferred that after entering multiple features simultaneously into the model, the risk associated with lower educational attainment was absorbed by other multidomain ADHD-related variables, thereby weakening the marginal effect of education itself [\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e]. Additionally, families with higher educational qualifications typically have better access to healthcare and diagnostic awareness, resulting in more children from these families receiving diagnoses [\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e]. Thus, the positive SHAP value associated with education is more likely reflective of diagnostic behaviour differences and covariate collinearity, rather than indicating that higher education itself increases ADHD risk.\u003c/p\u003e\u003cp\u003eFinally, this study systematically integrated the domains corresponding to the top ten features, which represent the majority of model contributions, and conducted cross-domain comparisons. To our knowledge, this is the first ADHD prediction study conducted within a behavioural framework to perform cross-domain group comparisons. Results indicated that the domain of early child cognitive and behavioural features contributed most significantly to the model. First, the features included in this domain were temporally closer to the diagnostic assessments at ages 5\u0026ndash;7, allowing the model to better capture the most discriminative symptomatic manifestations during this period [\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e]. This temporal proximity may have partly diluted the impact of demographic-domain features\u0026mdash;specifically mother\u0026rsquo;s academic qualification and household income quintiles\u0026mdash;in the ML model; at the same time, it may also have absorbed some of the explanatory power of the prenatal-domain features, thereby reducing that domain\u0026rsquo;s independent contribution.\u003c/p\u003e\u003cp\u003eSecond, previous studies have demonstrated that Conduct problems, Peer problems, and Emotional-Dysregulation serve as external indicators of restricted executive functioning in children and are significantly associated with subsequent ADHD symptoms [\u003cspan additionalcitationids=\"CR63\" citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e], whereas executive-function deficits are widely regarded as one of the core neurocognitive impairments in ADHD [\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e]. Therefore, this study hypothesises that early potential deficits in executive functioning may amplify ADHD risk through the expression of these specific features. However, this study did not include executive functioning features, and thus, the underlying mechanisms require further objective testing for validation.\u003c/p\u003e\u003cp\u003eNevertheless, prenatal factors and early maternal parenting style still jointly contributed 28.3% to the model's explanatory power, suggesting they carry independent information not covered by early child cognitive and behavioural development or demographic features. As these risk factors can be modified through public health interventions (e.g., smoking cessation during pregnancy, promotion of breastfeeding) and family interventions (e.g., parental behavioural management training) before ADHD symptoms emerge, they remain irreplaceably valuable for early screening and risk management.\u003c/p\u003e\u003cp\u003e\u003cb\u003elimitations\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe present study has several limitations. First, due to computational constraints, this study did not exhaust all possible algorithmic combinations to achieve optimal test-set performance\u0026mdash;for example, simultaneously incorporating undersampling, RFE, hyperparameter tuning, calibration, and threshold optimisation within a single pipeline to perform global optimisation. Such an approach would exponentially increase computational time due to multiple nested exploratory processes.\u003c/p\u003e\u003cp\u003eSecond, given the scarcity of ADHD cases among children, this study was unable to perform stratified analyses based on specific demographic variables. Instead, demographic variables were collectively incorporated into the ML models. Particularly regarding the Gender variable, the demographic descriptive analysis indicated a significant gender imbalance within the ADHD diagnostic group, potentially making Gender the most crucial feature in multiple models.\u003c/p\u003e\u003cp\u003eThird, to address training difficulties caused by extreme sample imbalance, RFE-based feature selection was conducted before model training. Although this improved model performance to some extent, it exacerbated the imbalance in the number of selected features across domains. For instance, the environmental exposure domain retained only one feature in the optimal XGBoost model, hindering meaningful cross-domain comparisons.\u003c/p\u003e\u003cp\u003eFourth, this study utilised only binary diagnostic labels from the MCS dataset as supervision signals, thereby limiting further exploration of specific core ADHD symptoms and the influence of corresponding clinical assessment scores within the machine learning framework. Although the MCS dataset contains SDQ Hyperactivity/Inattention (H/I) items with detailed scale scores, allowing for potential mapping onto core ADHD symptoms, the SDQ scale itself is not a clinical diagnostic instrument, and its results lack diagnostic authority for ADHD.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study developed an early childhood ADHD prediction model using the XGBoost machine learning algorithm. Its predictive performance was not only favourable compared to existing early childhood ADHD prediction models but also demonstrated the capacity to predict ADHD diagnosis at ages 5\u0026ndash;7 using behavioural-level data collected by age 3. More importantly, these findings are generalizable to the entire UK population. Although early child cognitive and behavioural features played a critical role in ADHD prediction, maternal prenatal factors and other postnatal domains should not be overlooked. In conclusion, this study offers valuable insights for future ADHD machine learning research, underscoring the importance of comprehensively considering the contributions of multidomain features to ADHD prediction.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e: The authors declare no competing interests\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval\u003c/strong\u003e: All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to participate:\u0026nbsp;\u003c/strong\u003eAll cohort members participating in this study provided informed consent signed by themselves or by their caregivers acting as proxies\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eZW: Contributed to the conception or design of the study, data analysis and interpretation, manuscript writing, and revision and review of the manuscript.HW: Contributed to partial manuscript writing, revision and review of the manuscript.RZ: Contributed to revision and review of the manuscript.LZ: Contributed to the conception or design of the study, revision and review of the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data used in this study were derived from the UK Millennium Cohort Study (MCS), which is publicly available for research purposes. The data on maternal prenatal and child postnatal variables from birth to age 7 that support the findings of this study have been deposited in the UK Data Service repository at: https://ukdataservice.ac.uk.The datasets involved in this study are classified as safeguarded level by the UK Data Service. Due to the sensitive nature of the data, raw data and individual-level identifiers cannot be publicly shared. Researchers wishing to access the MCS dataset must register with the UK Data Service and agree to the terms and conditions of data use.The analysis code has been made publicly available on GitHub: https://github.com/Anankysus/Early-ADHD-ML/tree/main. Additionally, detailed explanations of the variables/features used in the study are provided in the supplementary materials, to facilitate replication of the study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAmerican Psychiatric Association (2022) Diagnostic and statistical manual of mental disorders. Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR) 5:. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1176/appi.books.9780890425787\u003c/span\u003e\u003cspan address=\"10.1176/appi.books.9780890425787\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBrites C, Brites HD, de Almeida RP et al (2023) Early Diagnosis on ADHD. Is It Possible? Psychol 14:359\u0026ndash;370. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.4236/psych.2023.143021\u003c/span\u003e\u003cspan address=\"10.4236/psych.2023.143021\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZahmacioglu O, Kilic E (2017) Early diagnosis and treatment of ADHD are important for a secure transition to adolescence. Anatol J Psychiatry 18:1. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5455/apd.220054\u003c/span\u003e\u003cspan address=\"10.5455/apd.220054\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBoland H, DiSalvo M, Fried R et al (2020) A literature review and meta-analysis on the effects of ADHD medications on functional outcomes. J Psychiatr Res 123:21\u0026ndash;30. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jpsychires.2020.01.006\u003c/span\u003e\u003cspan address=\"10.1016/j.jpsychires.2020.01.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFord T, Goodman R, Meltzer H (2003) The British Child and Adolescent Mental Health Survey 1999: the Prevalence of DSM-IV Disorders. J Am Acad Child Adolesc Psychiatry 42:1203\u0026ndash;1211. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1097/00004583-200310000-00011\u003c/span\u003e\u003cspan address=\"10.1097/00004583-200310000-00011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePrasad V, West J, Kendrick D, Sayal K (2019) Attention-deficit/hyperactivity disorder: variation by socioeconomic deprivation. Arch Dis Child 104:802\u0026ndash;805. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1136/archdischild-2017-314470\u003c/span\u003e\u003cspan address=\"10.1136/archdischild-2017-314470\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e\u0026Oslash;vergaard KR, Oerbeck B, Friis S et al (2023) Early and repeated screening detects children with persistent attention-deficit/hyperactivity disorder. Eur Child Adolesc Psychiatry. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00787-023-02284-8\u003c/span\u003e\u003cspan address=\"10.1007/s00787-023-02284-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e\u0026Oslash;vergaard KR, Oerbeck B, Friis S et al (2018) Attention-Deficit/Hyperactivity Disorder in Preschoolers: The Accuracy of a Short Screener. J Am Acad Child Adolesc Psychiatry 57:428\u0026ndash;435. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jaac.2018.03.008\u003c/span\u003e\u003cspan address=\"10.1016/j.jaac.2018.03.008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChaulagain A, Lyhmann I, Halm\u0026oslash;y A et al (2023) A systematic meta-review of systematic reviews on attention deficit hyperactivity disorder. Eur Psychiatry 66. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1192/j.eurpsy.2023.2451\u003c/span\u003e\u003cspan address=\"10.1192/j.eurpsy.2023.2451\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi L, Lagerberg T, Chang Z et al (2020) Maternal pre-pregnancy overweight/obesity and the risk of attention-deficit/hyperactivity disorder in offspring: a systematic review, meta-analysis and quasi-experimental family-based study. Int J Epidemiol 49:857\u0026ndash;875. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/ije/dyaa040\u003c/span\u003e\u003cspan address=\"10.1093/ije/dyaa040\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDong T, Hu W, Zhou X et al (2018) Prenatal exposure to maternal smoking during pregnancy and attention-deficit/hyperactivity disorder in offspring: A meta-analysis. Reprod Toxicol 76:63\u0026ndash;70. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.reprotox.2017.12.010\u003c/span\u003e\u003cspan address=\"10.1016/j.reprotox.2017.12.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSerati M, Barkin JL, Orsenigo G et al (2017) Research Review: The role of obstetric and neonatal complications in childhood attention deficit and hyperactivity disorder \u0026ndash; a systematic review. Child Psychol Psychiatry 58:1290\u0026ndash;1300. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/jcpp.12779\u003c/span\u003e\u003cspan address=\"10.1111/jcpp.12779\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e(2017) Attention deficit/hyperactivity disorder and risk of injuries: a systematic review and meta-analysis. J Inj Violence Res 9:. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5249/jivr.v9i2.858\u003c/span\u003e\u003cspan address=\"10.5249/jivr.v9i2.858\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang A, Wu K, Cai Z et al (2021) Association between postnatal second-hand smoke exposure and ADHD in children: a systematic review and meta-analysis. Environ Sci Pollut Res 28:1370\u0026ndash;1380. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11356-020-11269-y\u003c/span\u003e\u003cspan address=\"10.1007/s11356-020-11269-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTseng P-T, Yen C-F, Chen Y-W et al (2019) Maternal breastfeeding and attention-deficit/hyperactivity disorder in children: a meta-analysis. Eur Child Adolesc Psychiatry 28:19\u0026ndash;30. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00787-018-1182-4\u003c/span\u003e\u003cspan address=\"10.1007/s00787-018-1182-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSalazar G, Iniesta R, Bellato A et al (2024) Individualized prediction models in ADHD: a systematic review and meta-regression. Mol Psychiatry. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41380-024-02606-5\u003c/span\u003e\u003cspan address=\"10.1038/s41380-024-02606-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDumitrescu E, Hu\u0026eacute; S, Hurlin C, Tokpavi S (2021) Machine Learning for Credit Scoring: Improving Logistic Regression with Non-Linear Decision-Tree Effects. Eur J Oper Res. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ejor.2021.06.053\u003c/span\u003e\u003cspan address=\"10.1016/j.ejor.2021.06.053\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 297:\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFokkema M, Iliescu D, Greiff S, Ziegler M (2022) Machine Learning and Prediction in Psychological Assessment. Eur J Psychol Assess 38:165\u0026ndash;175. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1027/1015-5759/a000714\u003c/span\u003e\u003cspan address=\"10.1027/1015-5759/a000714\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu YS, Talarico F, Metes D et al (2024) Early identification of children with Attention-Deficit/Hyperactivity Disorder (ADHD). PLOS Digit Health 3:e0000620\u0026ndash;e0000620. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pdig.0000620\u003c/span\u003e\u003cspan address=\"10.1371/journal.pdig.0000620\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGao T, Yang L, Zhou J et al (2024) Development and validation of a nomogram prediction model for ADHD in children based on individual, family, and social factors. J Affect Disord. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jad.2024.04.069\u003c/span\u003e\u003cspan address=\"10.1016/j.jad.2024.04.069\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGarcia-Argibay M, Zhang-James Y, Cortese S et al (2022) Predicting childhood and adolescent attention-deficit/hyperactivity disorder onset: a nationwide deep learning approach. Mol Psychiatry. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41380-022-01918-8\u003c/span\u003e\u003cspan address=\"10.1038/s41380-022-01918-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTer-Minassian L, Viani N, Wickersham A et al (2022) Assessing machine learning for fair prediction of ADHD in school pupils using a retrospective cohort study of linked education and healthcare data. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1136/bmjopen-2021-058058\u003c/span\u003e\u003cspan address=\"10.1136/bmjopen-2021-058058\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. BMJ Open 12:\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNurisa M, Dastghaibyfard G, Hadianfard H (2022) Predicting ADHD from early childhood data using data mining. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.researchgate.net/publication/360185624_Predicting_ADHD_from_early_childhood_data_using_data_mining\u003c/span\u003e\u003cspan address=\"https://www.researchgate.net/publication/360185624_Predicting_ADHD_from_early_childhood_data_using_data_mining\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eConnelly R, Platt L (2014) Cohort Profile: UK Millennium Cohort Study (MCS). Int J Epidemiol 43:1719\u0026ndash;1725. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/ije/dyu001\u003c/span\u003e\u003cspan address=\"10.1093/ije/dyu001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShepherd P, Gilbert E (2019) Millennium Cohort Study Ethical Review and Consent. London: Centre for Longitudinal Studies, UCL Institute of Education. Retrieved from\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMin X, Li C, Yan Y (2021) Parental Age and the Risk of ADHD in Offspring: A Systematic Review and Meta-Analysis. Int J Environ Res Public Health 18:4939. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/ijerph18094939\u003c/span\u003e\u003cspan address=\"10.3390/ijerph18094939\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWilson SV, Cebere B, Myatt J, Wilson S (2022) miceforest (Version 0.6.3). Zenodo. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.742863\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.742863\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCentre for Longitudinal Studies (2020) Millennium Cohort Study User Guide (Surveys 1\u0026ndash;5) (9th Edition). London: UCL Institute of Education\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiaw A, Wiener M (2002) Classification and Regression by randomForest. Retrieved R News 2 18\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen T, Guestrin C (2016) XGBoost: a Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD \u0026rsquo;16 1:785\u0026ndash;794. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/2939672.2939785\u003c/span\u003e\u003cspan address=\"10.1145/2939672.2939785\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKe G, Meng Q, Finley T et al (2017) LightGBM: a highly efficient gradient boosting decision tree. Curran Associates Inc., Red Hook, NY, USA, pp 3149\u0026ndash;3157\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20:273\u0026ndash;297. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/BF00994018\u003c/span\u003e\u003cspan address=\"10.1007/BF00994018\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePeterson L (2009) K-nearest neighbor. Scholarpedia 4:1883. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.4249/scholarpedia.1883\u003c/span\u003e\u003cspan address=\"10.4249/scholarpedia.1883\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRumelhart DE, Hinton GE, Williams RJ (1986) Learning representations by back-propagating errors. Nature 323:533\u0026ndash;536. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/323533a0\u003c/span\u003e\u003cspan address=\"10.1038/323533a0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKetende S, Jones E (2011) The Millenium Cohort Study User Guide to Analysing. MCS Data Using STATA\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePedregosa F, Varoquaux G, Gramfort A et al (2011) Scikit-learn: Machine Learning in Python. J Mach Learn Res 12. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5555/1953048.2078195\u003c/span\u003e\u003cspan address=\"10.5555/1953048.2078195\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe H, Garcia EA (2009) Learning from Imbalanced Data. IEEE Trans Knowl Data Eng 21:1263\u0026ndash;1284. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/tkde.2008.239\u003c/span\u003e\u003cspan address=\"10.1109/tkde.2008.239\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYouden WJ (1950) Index for rating diagnostic tests. Cancer 3:32\u0026ndash;35. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/1097-0142(1950)3:1\u0026lt;32::aid-cncr2820030106\u0026gt;3.0.co;2-3\u003c/span\u003e\u003cspan address=\"10.1002/1097-0142(1950)3:1%3C32::aid-cncr2820030106%3E3.0.co;2-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2002) SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res 16:321\u0026ndash;357. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1613/jair.953\u003c/span\u003e\u003cspan address=\"10.1613/jair.953\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWollenhaupt-Aguiar B, Librenza-Garcia D, Giovana, Bristot et al (2019) Differential biomarker signatures in unipolar and bipolar depression: A machine learning approach. Psychiatry 54:393\u0026ndash;401. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/0004867419888027\u003c/span\u003e\u003cspan address=\"10.1177/0004867419888027\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLundberg S, Lee S-I (2017) A Unified Approach to Interpreting Model Predictions. In: arXiv.org. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/abs/1705.07874v2\u003c/span\u003e\u003cspan address=\"https://arxiv.org/abs/1705.07874v2\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang D, Thun\u0026eacute;ll S, Lindberg U et al (2022) Towards better process management in wastewater treatment plants: Process analytics based on SHAP values for tree-based machine learning methods. J Environ Manage 301:113941. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jenvman.2021.113941\u003c/span\u003e\u003cspan address=\"10.1016/j.jenvman.2021.113941\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSpencer NJ, Ludvigsson J, Bai G et al (2022) Social gradients in ADHD by household income and maternal education exposure during early childhood: Findings from birth cohort studies across six countries. PLoS ONE 17:e0264709. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0264709\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0264709\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRabinowicz A, Rosset S (2021) Trees-Based Models for Correlated Data. arXiv (Cornell University). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arxiv.2102.08114\u003c/span\u003e\u003cspan address=\"10.48550/arxiv.2102.08114\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuyon I, Elisseeff A (2003) An introduction to variable and feature selection. J Mach Learn Res. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5555/944919.944968\u003c/span\u003e\u003cspan address=\"10.5555/944919.944968\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGoodman R (1997) The Strengths and Difficulties Questionnaire: A Research Note. J Child Psychol Psychiatry 38:581\u0026ndash;586. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1469-7610.1997.tb01545.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1469-7610.1997.tb01545.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIizuka C, Yamashita Y, Nagamitsu S et al (2010) Comparison of the strengths and difficulties questionnaire (SDQ) scores between children with high-functioning autism spectrum disorder (HFASD) and attention-deficit/hyperactivity disorder (AD/HD). Brain Develop 32:609\u0026ndash;612. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.braindev.2009.09.009\u003c/span\u003e\u003cspan address=\"10.1016/j.braindev.2009.09.009\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCroft S, Stride C, Maughan B, Rowe R (2015) Validity of the Strengths and Difficulties Questionnaire in Preschool-Aged Children. Pediatrics 135:e1210\u0026ndash;e1219. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1542/peds.2014-2920\u003c/span\u003e\u003cspan address=\"10.1542/peds.2014-2920\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSammons P, Sylva S, Melhuish E et al (2004) The Effective Provision of Pre-school Education (EPPE) Project: Technical Paper 11: The continuing effect of pre-school education at age 7 years. Institute of Education, London\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMelhuish E, Hanna K, Quinn L et al (2004) The Effective Pre-school Provision in Northern Ireland Project, Technical Paper 11: Pre-school Experience and Social/Behavioural Development At the End of Year 3 of Primary School. Stranmillis University, Belfast, N.I.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEfron D, Nicholson JM, Anderson V et al (2020) ADHD at Age 7 and Functional Impairments at Age 10. Pediatrics 146:e20201061. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1542/peds.2020-1061\u003c/span\u003e\u003cspan address=\"10.1542/peds.2020-1061\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShaw P, Stringaris A, Nigg J, Leibenluft E (2014) Emotion Dysregulation in Attention Deficit Hyperactivity Disorder. Am J Psychiatry 171:276\u0026ndash;293. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1176/appi.ajp.2013.13070966\u003c/span\u003e\u003cspan address=\"10.1176/appi.ajp.2013.13070966\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAntony EM-A, Pihlajam\u0026auml;ki MR, Speyer LG, Murray AL (2021) Does emotion dysregulation mediate the association between ADHD symptoms and internalizing problems? A longitudinal within-person analysis in a large population-representative study. Edinburgh Research Explorer (University of Edinburgh. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.31234/osf.io/32dky\u003c/span\u003e\u003cspan address=\"10.31234/osf.io/32dky\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGoodman R (2015) Scoring the Strengths \u0026amp; Difficulties Questionnaire for 2\u0026ndash;4 year olds\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHartman CA, Luteijn E, Serra M, Minderaa R (2006) Refinement of the Children\u0026rsquo;s Social Behavior Questionnaire (CSBQ): An Instrument that Describes the Diverse Problems Seen in Milder Forms of PDD. J Autism Dev Disord 36:325\u0026ndash;342. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10803-005-0072-z\u003c/span\u003e\u003cspan address=\"10.1007/s10803-005-0072-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLundberg SM, Erion G, Chen H et al (2020) From local explanations to global understanding with explainable AI for trees. Nat Mach Intell 2:56\u0026ndash;67. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s42256-019-0138-9\u003c/span\u003e\u003cspan address=\"10.1038/s42256-019-0138-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eClaussen AH, Holbrook JR, Hutchins HJ et al (2022) All in the Family? A Systematic Review and Meta-analysis of Parenting and Family Environment as Risk Factors for Attention-Deficit/Hyperactivity Disorder (ADHD) in Children. Prev Sci 25. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11121-022-01358-4\u003c/span\u003e\u003cspan address=\"10.1007/s11121-022-01358-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMohammadian M, Khachatryan LG, Vadiyan FV et al (2025) The association between maternal tobacco smoking during pregnancy and the risk of attention-deficit/hyperactivity disorder (ADHD) in offspring: A systematic review and meta-analysis. PLoS ONE 20:e0317112\u0026ndash;e0317112. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0317112\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0317112\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGal\u0026eacute;ra C (2011) Early Risk Factors for Hyperactivity-Impulsivity and Inattention Trajectories From Age 17 Months to 8 Years. Arch Gen Psychiatry 68:1267. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1001/archgenpsychiatry.2011.138\u003c/span\u003e\u003cspan address=\"10.1001/archgenpsychiatry.2011.138\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOwens J (2021) Parental intervention in school, academic pressure, and childhood diagnoses of ADHD. Soc Sci Med 272:113746. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.socscimed.2021.113746\u003c/span\u003e\u003cspan address=\"10.1016/j.socscimed.2021.113746\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSonuga-Barke EJS, Halperin JM (2010) Developmental phenotypes and causal pathways in attention deficit/hyperactivity disorder: potential targets for early intervention? Child Psychol Psychiatry 51:368\u0026ndash;389. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1469-7610.2009.02195.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1469-7610.2009.02195.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang Y, Shields GS, Zhang Y et al (2022) Child executive function and future externalizing and internalizing problems: A meta-analysis of prospective longitudinal studies. Clin Psychol Rev 97:102194. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cpr.2022.102194\u003c/span\u003e\u003cspan address=\"10.1016/j.cpr.2022.102194\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMarques S, Correia-de-S\u0026aacute; T, Guardiano M et al (2024) Emotion dysregulation and depressive symptoms mediate the association between inhibitory control difficulties and aggressive behaviour in children with ADHD. Front Psychiatry 15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fpsyt.2024.1329401\u003c/span\u003e\u003cspan address=\"10.3389/fpsyt.2024.1329401\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLan Y-T, Liu X-P, Fang H-S (2018) Randomized control study of the effects of executive function training on peer difficulties of children with attention-deficit/hyperactivity disorder C subtype. Appl Neuropsychology: Child 9:41\u0026ndash;55. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/21622965.2018.1509003\u003c/span\u003e\u003cspan address=\"10.1080/21622965.2018.1509003\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWillcutt EG, Doyle AE, Nigg JT et al (2005) Validity of the Executive Function Theory of Attention-Deficit/Hyperactivity Disorder: A Meta-Analytic Review. Biol Psychiatry 57:1336\u0026ndash;1346. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.biopsych.2005.02.006\u003c/span\u003e\u003cspan address=\"10.1016/j.biopsych.2005.02.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"ADHD, Early prediction, Prenatal, Postnatal, Machine learning","lastPublishedDoi":"10.21203/rs.3.rs-7134745/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7134745/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLimited attention has been devoted to developing machine-learning models that use behavioural data for the early prediction of childhood attention-deficit/hyperactivity disorder (ADHD), particularly in the United Kingdom. Therefore, this study evaluated the predictive performance of six machine learning approaches in a cohort of 9,385 children (259 with ADHD, 9,126 controls) from the UK Millennium Cohort Study. After selecting the optimal model, we comprehensively compared the relative contributions of prenatal and postnatal (0\u0026ndash;3 years) multi-domain features to its predictive performance. Results indicated that XGBoost achieved the highest performance on the test set (AUC\u0026thinsp;=\u0026thinsp;0.881), effectively balancing the rates of false positives and false negatives. Specifically, \"Conduct problems\" is the most significant predictor across all models. Among postnatal features, early childhood cognitive and behavioural development represented the most influential domain, contributing approximately 51.9% SHAP value to the model. Nonetheless, other domain features (e.g. prenatal features) show non-negligible contributions. By establishing robust predictive performance, this research addresses an existing gap in machine learning-based studies of childhood ADHD within the UK context. Furthermore, as the first study to quantitatively evaluate the contribution of multiple behavioural domain features to predictive model performance in ADHD, this work provides valuable insights for future model development.\u003c/p\u003e","manuscriptTitle":"Early prediction of childhood ADHD using prenatal and early postnatal behavioural features: evaluation across six machine-learning algorithms","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-28 07:24:01","doi":"10.21203/rs.3.rs-7134745/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"2565c16a-9b78-4c7b-8389-b8059c195f99","owner":[],"postedDate":"July 28th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-08-13T07:09:48+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-28 07:24:01","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7134745","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7134745","identity":"rs-7134745","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}