Explainable machine learning of PROGRESS-Plus social factors predicts cognitive trajectories after traumatic brain injury

Explainable machine learning of PROGRESS-Plus social factors predicts cognitive trajectories after traumatic brain injury | 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 Article Explainable machine learning of PROGRESS-Plus social factors predicts cognitive trajectories after traumatic brain injury Jingwen Xu, Urooba Shaikh, Thaisa Tylinski Sant’Ana, Tatyana Mollayeva This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8585532/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 30 Mar, 2026 Read the published version in Scientific Reports → Version 1 posted You are reading this latest preprint version Abstract Scientific research on social parameters for prognosis after traumatic brain injury (TBI) is evolving, yet results remain heterogeneous, and predictors for risk stratification are lacking. To understand how social parameters are linked to cognitive outcomes after TBI, we applied machine learning (ML) algorithms using data from 30 published studies including 2,364 participants with TBI (72% male; 55% mild, 45% moderate-severe injury). We extracted and harmonised longitudinal data following the PROGRESS-Plus framework, and used the data as predictors of rate of change in cognition post-TBI. We developed random forest, gradient boosting (GB), and extreme GB predictive models, accounting for time from injury to baseline assessment, time between assessments, and country-level structural indicators. We used mean absolute error and root mean squared error to evaluate model performance, and Shapley Additive Explanations analysis for explanatory model predictions. Results highlighted time interval, country-level structural indicators, age, and variation in education as key predictors for rate of change for both injury severities. Sensitivity analyses for predicting rate of change in executive function and learning and memory confirmed the robustness of the results. Our work contributes to novel ML research for understanding prognosis and advancing precision in predicting cognitive outcomes after TBI. Health sciences/Medical research Health sciences/Neurology Biological sciences/Neuroscience Artificial intelligence Common data elements Data harmonization Evidence gap Information management Social determinants of health Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Research on the course and prognosis of cognitive function following traumatic brain injury (TBI), which refers to structural and/or physiological disruption of brain function due to an external force 1 , is rapidly developing but remains characterized by extensive inter-patient and intra-patient heterogeneity. Adverse outcomes are attributed to greater injury severity and advanced age, highlighting the challenges in care and poor prognosis for these patients 2 , 3 . Recently, the view of TBI of mild severity is progressing through evidence towards a condition with long-term consequences for cognition 4 , 5 rather than a singular event. Furthermore, it has been proposed that TBI should not be seen as a random event and accident, but rather as a health status transition on a continuum, where the magnitude of social adversities and comorbidities increase the risk of injury severity, its adverse course, and expedited 2 , 6 – 9 cognitive decline, across injury severities 10 , 11 . With this in mind, recent research has started to increasingly pose social equity hypotheses and investigate the role of social parameters in recovery course and long-term outcomes. However, the challenge remains that social parameters are intertwined with injury-related and environmental risks, age-related parameters, and broader social adversities, forming a vastly complex network of associations 11 , 12 . Capturing and disentangling these associations on a time continuum presents a substantial challenge for traditional hypothesis-driven research, which typically assumes linear relationships and examines a limited number of variables 13 , 14 . As a result, important sources of variability in TBI prognosis remain insufficiently characterised in scientific research, creating a knowledge gap in research, policy, and practice 14 . Emerging advances in computational and machine learning (ML) approaches offer a timely opportunity to address this gap by enabling more systematic evaluation of variable importance and the complex interplay of social and demographic parameters in shaping cognitive trajectories after TBI. One method to increase confidence in the prognosis of cognitive course after TBI is to apply computational and ML approaches to longitudinal data from published TBI cohorts. These approaches enable the evaluation of variable importance and the complex interplay of social and demographic parameters shaping cognitive outcomes after TBI. Here, we present a new analysis of data from 30 published longitudinal cohort studies characterising change in cognitive test performance with 2,364 participants with TBI (72% male, mean age 32 years; 55% mild, 45% moderate/severe TBI) 15 . We leveraged advancements in computational and ML approaches, with the aim to: (1) develop and validate predictive models of longitudinal cognitive change after TBI using harmonised social, demographic, and injury-related data from published cohort studies; (2) compare the predictive performance of multiple ML approaches (random forest, gradient boosting, and extreme gradient boosting) in modelling standardised rates of cognitive change following TBI, stratified by injury severity, and (3) identify and interpret the relative importance of social and demographic parameters in predicting cognitive trajectories after TBI. Our overarching aim was to determine the prognostic significance of social risk stratification in TBI by leveraging an interpretable ML framework. Through this approach, we sought to move beyond traditional injury-related predictors and elucidate how social profiles of study samples shape longitudinal cognitive course after TBI. This approach to characterizing heterogeneity in cognitive outcomes will equip researchers and policymakers with actionable evidence for understanding and interpreting heterogeneity in cognitive outcomes, and informing more equitable, data-driven prediction. Results We included 30 published studies 16–45 and 43 cohorts within them (i.e., Mandleberg et al. provided eight cohorts 30 , Dikmen et al. provided three cohorts 19 , and Kontos et al. 24 , Macciochi et al. 28 , and Covassin et al. 18 provided two cohorts, each), which provided a total of 218 cognitive outcome data points (see dataset). The analysis included 218 data points collected on at least two time points after injury comprising 2,364 adults (72% male) with TBI. Of these, 57% corresponded to data of participants with mild TBI, with the remaining data points representing moderate (4.6%), moderate-severe TBI (24%) and severe TBI (16%). To mitigate the influence of imbalance in our data in injury severity categories and to preserve clinically meaningful distinctions, moderate, moderate–severe, and severe injuries were combined into a single ‘moderate–severe’ category, and models were trained separately for mild and moderate–severe TBI. Published data spanned five decades, with the majority (87%) published after the year 2000. All studies were conducted in high- and middle-income countries. Most cohorts originated from the USA (53%), followed by New Zealand and UK (9.6% each), China (6.9%), Norway (5.0%), South Korea (4.6%), Brazil (4.1%), and Canada and Australia (3.7%, each) (Table 1). To strengthen the use of country of study origin as a proxy for place of residence, we used country of English language dominance instead of country of study origin in the modelling process, categorizing it under place of residence. Insert Table 1 here Data extraction process After combining moderate and severe injury severity with moderate-severe data points, the final dataset resulted in 121 data points for mild, and 97 data points for moderate-severe TBI (Supplement Table 1). Published data spanned five decades, with the majority (87%) published after the year 2000. Data points represented nine countries, all of high- and middle-income, across global regions. Most cohorts originated from the USA (53%), followed by New Zealand and UK (9.6% each), China (6.9%), Norway (5.0%), South Korea (4.6%), Brazil (4.1%), and Canada and Australia (3.7%, each) (Table 1). To strengthen the use of country of study origin as a proxy for place of residence, we used country of English language dominance instead of country of study origin in the modelling process, categorizing it under place of residence. Reporting of study sample characteristics varied across cohorts, regardless of country of study origin. Gender/sex composition of study samples was reported in all cohorts, and after harmonisation proportion of males in samples ranged from 0.44 42 to 1 16,24 (Table 1). Cohorts’ education was reported using a variety of formats. After harmonisation to years of formal education, mean education levels ranged from 7.1 17,37 to 14.7 42 years across cohorts; 11 25,26 to 14.7 42 years for mild, and 7.1 17,37 to 13.17 20 years for moderate-severe TBI cohorts. Education SDs ranged from 1.11 35 to 4 17,37 years, 1.5 28 to 4 17,37 years for mild, and 1.11 35 to 3.3 39 years for moderate-severe TBI cohorts. Other PROGRESS-Plus variables (i.e., race/ethnicity, occupation, social capital, and socioeconomic status) were represented by a limited number of data points and were not used in data analysis (Supplement Table 1). Of the Plus parameters, age and age SD were available for all data points. The pooled mean age across studies ranged from 18.8 16 to 43.68 40 years, 18.8 16 to 43.68 40 for mild TBI and 25.2 36 to 41 33 years for moderate-severe TBI (please see dataset for specifics and data by injury severity). The baseline assessment time (time zero) ranged from ~12 hours 16 to ~10 years post-injury 22 and final follow-up assessments ranged from five days 18 to ~20 years 22 post-TBI (Supplement table 1), with the mean number of days from injury to baseline assessment being 11.2 for mild and 348 for moderate-severe TBI cohorts. For the purpose of observing rate of change per month as the outcome, we then converted time-related variables to the common metric of months by dividing by 30.4, the average number of days in a month (Supplement Table 1). The distribution of data points across cognitive test domains, overall and by injury severity, was as follows, from highest to lowest: executive function (118; 55% mild), learning and memory (105, 49% mild), perceptual-motor (87, 64% mild), information processing speed (85, 72% mild), complex attention (84, 60% mild), language (27, 41% mild), and social cognition (6, 100% mild). We refer the reader to Supplement Table 1 for data used in this synthesis. The cognitive outcome data for ML was a nominal variable across domains of cognition, standardised as rate of change per month (see Methods section).The rate of change was also calculated for domains of cognition with sufficient data: executive function and learning and memory. The outcome data points were collected at baseline and last follow-up, and formed the foundation for the predictive modelling of cognitive trajectories presented in this study (Supplement Table 1). From baseline to last follow-up assessment, the median standardised rate of change per month was 0.233 for mild, and 0.112 for moderate-severe TBI. We refer the reader to the figures for results of the standardised rate of change for mild and moderate-severe TBI for each of the PROGRESS-Plus parameters with available data: the country of study origin, gender/sex and education overall, and by domains of cognition (Supplement Figure 1). Data preprocessing We created heat maps with correlation coefficient matrices for mild and moderate-severe TBI to identify associations among PROGRESS-Plus variables (Figure 2 and Supplement Figure 2 for country-specific associations). We treated all PROGRESS-Plus variables with complete data coverage as primary predictors and included them in the analyses irrespective of the statistical significance of their association with the rate of change. While country of study origin was available for each cohort, in order to use it efficiently (i.e., considering historical evolution) we used Gender Inequality Index (GII) and Human Development Index (HDI) as contextual measures and effect modifiers in the modelling process. These country-level structural indicators contributed to measurable heterogeneity across mild and moderate-severe TBI cohorts in several key predictors (Figure 2). Assigned HDI and GII values ranged from 0.689 25 to 0.928 24,32,34,42 and 0.009 41 to 0.269 38 , respectively. Temporal trends indicated a gradual change in values in cohorts coming from the same country (refer to dataset). Along with PROGRESS-Plus parameters, we also evaluated whether time-related variables (i.e., time from injury to baseline assessment and time between baseline and last follow up assessment) were associated with the rate of change across injury severity cohorts (Figure 3). Time from injury to baseline assessment and time between baseline assessment and last follow-up varied (Table 2) and were negatively and positively associated with the rate of change in mild and moderate-severe TBI respectively. We retained these variables as covariates in models. Insert Figure 2 Machine learning modeling Results of three supervised machine-learning algorithms, Random Forest (RF), Gradient Boosting (GB) and extreme GB (XGBoost), including both overall cognitive rate of change and domain-specific outcomes, respectively, can be seen in Figures 3, 5 and 7, for mild and moderate-severe TBI. Feature importance heatmaps of each of the models are presented in Figure 3a and b, for mild and moderate-severe TBI, respectively. For mild TBI, across the three models, the top PROGRESS-Plus feature was age, followed by education SD and the contextual parameter of GII. For moderate-severe TBI cohorts, the top PROGRESS-Plus feature was age SD and time related variables. Insert Figure 3 Model evaluation and interpretation Model performance metrics are summarised in Table 2. Across all algorithms, predictive accuracy was comparable, with XGBoost slightly better in mean absolute error (MAE), and RF slightly better in root mean squared error (RMSE) in mild and moderate-severe TBI cohorts (Table 2) Insert Table 2 The pattern of tuned hyperparameters (Table 3) across algorithms suggested different levels of structure in data in both mild and moderate-severe TBI. We observed that RF for both injury severity strata captured very shallow trees (i.e., maximum depth = 2) with the complexity in predictive signals explained by low-order interactions. The GB model favoured deeper individual trees (i.e., maximum depth of 9 and 6 for mild and moderate-severe TBI, respectively) where each successive tree focused on residual errors, allowing more complex interaction within a single tree compared to RF. The mild TBI model highlighted a very low learning rate (i.e., 0.01), with deeper trees and more iterations. The moderate-severe model used a higher learning rate (i.e., 0.1) with fewer trees and shallower depths. The XGBoost model, across mild and moderate-sever TBI strata, selected deeper trees using additional regularization to tolerate deeper depths without overfitting. The learning rate remained small, and overall complexity was determined by interactions with a number of boosting runs (Table 3). To support interpretation beyond hyperparameters, we examined variable influence and marginal effects (e.g., partial dependence/SHAP) to understand how each algorithm captured non-linear relationships and interactions in the data (Figure 4). Insert Table 3 For prediction, the Shapley Additive Explanations (SHAP) ranking showed that across models, time interval, GII, education (picked by RF), and age emerged as the strongest predictors of standardised rate of change for mild TBI. For moderate-severe TBI cohorts, across models, baseline assessment, time interval, age, and GII (picked by RF), emerged as the strongest predictors of standardised rate of change (Figure 4). Insert Figure 4 Internal validation and sensitivity analyses To ensure the robustness of the results, we performed 5-fold cross-validation to verify generalisability across folds (Table 4). For both injury severities, XGBoost showed slightly superior performance for standardised rate of change across data points. The effect of all the primary predictors, covariates, and the effect modifiers remained the same (Supplement figures 4 and 5). Insert Table 4 Further analyses using a subset of data on executive function and learning and memory domains of cognition, which had relatively balanced data between mild and moderate-severe TBI cohorts, confirmed our hypothesis (i.e., that PROGRESS-Plus parameters would exert a more pronounced effect on domain-specific cognitive outcomes than on overall cognition, stronger for mild than moderate-severe TBI). Results showed that all the values of relative contributions for PROGRESS-Plus and time variables varied across models, however they remained largely the same as in the main analyses (Supplement Figures 6 and 10 for executive function and learning and memory TBI cohorts, respectively). The top three features for prediction of rate of change in executive function domain in mild TBI cohorts identified by the SHAP analyses were age, time interval, and education SD. These results were similar to the main analysis, with the exception of GII whose predictive value was reduced compared to the main analysis. In moderate-severe TBI cohorts, the top three predictive features remained as time interval, time from injury to baseline assessment, and age SD (Supplement Figure 7). The top three consistent features for prediction of rate of change in the learning and memory domain in mild TBI cohorts identified by the SHAP analyses were education SD, time interval, and GII. These results were similar to the main analysis, with the exception of age whose predictive value was reduced compared to the main analysis. In moderate-severe TBI cohorts, the top three predictive features remained as time interval, time from injury to baseline assessment, and age SD, similar to the main analysis (Supplement Figure 11). Model performance for both domains remained similar to that of the main analysis (Supplement Table 4). The results of sensitivity analyses using 5-fold cross-validation as an internal validation approach to evaluate the robustness and generalizability of model estimates for specific domains of cognition are presented in Supplement Figures 8-9 and 12-13. The targeted elimination of multicollinear age and education SD parameters provided further validity regarding the impact of time interval, age, and GII in mild TBI cohorts. In moderate-severe TBI, time related variables were further validated (Supplement Figures 14 and 15). Discussion In this study, we report innovative explainable ML research using published longitudinal data on the course of cognition after TBI, investigating PROGRESS-Plus characteristics of study samples as predictors of rate of change in cognition after injury. The results support the importance of considering social parameters in post-injury outcomes, and the ability of ML to explain heterogeneity 46 in the course of cognition overall and by cognitive domain following TBI of equal severity. Our study has research, clinical, and policy implications. We found that among the most important predictors of cognitive change after TBI in published research across domains and injury severities were age, time-related parameters (i.e., time from injury to baseline assessment and time interval from baseline to last follow-up assessment), and country-level structural indicators. Given the pronounced variation in age and age SD across cohorts, the discussion brings attention to how researchers deal with age in their analyses in longitudinal studies. This is true for both mild and moderate-severe TBI. Age-related deterioration affects all people, and is reflected across domains of cognition, reported for both mild and moderate-severe TBI cohorts 1 , 47 , 48 . The mechanisms that mediate age-related processes and changes reflected in the test performance of cohorts with TBI included in this study are not entirely clear; however, these processes are likely to be influenced by injury severity (Supplement Fig. 1 b, e, f). Our results suggest that aging processes may be accelerated to a greater extent after moderate–severe TBI as compared to mild, highlighting that neurorecovery and neurodegeneration occur simultaneously and therefore the interval between assessments does not carry the same meaning across injury severity cohorts. We found a higher rate of change in cognitive outcomes in moderate–severe TBI (0.23) than in mild TBI (0.11), highlighting the value of how machine learning in capturing temporal effects in predicting outcomes. Our results also align with the random damage theories 49 , 50 , which situate around the disturbed balance between ongoing damage and repair that occurs in the process of natural aging, reflected in cognitive test performance. When the process is disturbed by brain injury, the balance of ongoing damage and repair is further dysregulated, and therefore was picked by ML as key predictors of the rate of change, to a greater extent in moderate-severe TBI than mild (Supplement Fig. 1 b, e, f). The age of research participants, and the SD of the cohort, are, therefore expected to be seen, as it would affect the processes of both natural ageing and constrict recovery after injury, which is not expected to be uniform in participants of various ages. In addition, participants of different age in the cohort, reflected in the age SD of the cohort, are also more or less likely to suffer from multiple diseases associated with aging, which were rarely reported on in published research we included in the dataset, and which we were not able to test as predictors of the rate of change. Several longitudinal studies have uncovered that the age-specific strain caused by the onset of TBI in the presence of comorbidity is associated with cognitive outcomes, regardless of TBI severity 1 , 6 , 9 . Future research should investigate age and age-related effects precisely to understand their predictive role in prognosis. Time-related variables (i.e., time from injury to baseline assessment and time interval from baseline to last follow-up assessment, Supplement Table 2 ), emerged as important predictors of rate of change, and therefore emphasize the critical role of time in prognostic research. We observed that the rate of cognitive change after TBI differed between mild and moderate-severe cohorts (Table 1 ). This finding aligns with prior research and with current evidence defining severe TBI as a risk factor for long-term cognitive decline 51 , 52 . This study provides evidence that time affects rate of change across executive function and learning and memory domains of cognition, but we did not have sufficient data to test the effect on other cognitive domains (i.e., language, perceptual motor, complex attention, information processing speed, social cognition). Because our prior systematic review 1 , 53 and observational studies 34 , 54 brought attention to the fact that the course of cognition and recovery were not uniform and were dependent on the baseline assessment, future research should consider the implications of time after injury and baseline assessment in prognosis of cognitive domain-specific risks following injury. Our results on relevance of baseline assessments (i.e., time zero) in prognosis were especially pronounced in moderate-severe TBI cohorts, with the mean baseline assessment conducted at around one year post injury, with great heterogeneity between cohorts (Table 1 and dataset). In mild TBI, the mean baseline assessment was less than two weeks, with more homogeneity among samples (Table 1 ). At the point of baseline assessment, many participants in the mild TBI cohort may have reached a recovery plateau, which may not be the case for moderate-severe TBI, and therefore the time effects were not as pronounced in ML prediction for mild as in moderate-severe cohorts. Our results in sensitivity analyses confirmed the robustness of time effects, which underscore a critical need for coordinated research and policy efforts that explicitly integrate discussion on timing of research concerning prognosis based on injury severity, as this would impact scientific evidence with policy and practice implications. This is especially important as in both our prior hypothesis-driven approaches, we faced challenges in characterizing these temporal effects. Given that recovery occurs in parallel with processes of aging, these processes may reinforce one another with differing speed based on time that emerged from injury. Our results underscore that time is not merely a methodological detail, but a determinant of predictive accuracy and scientific interpretation. Future research should consider the benefits of ML for the prediction of outcomes after TBI with greater precision to time. Country-level indicators also warrant discussion. We systematically integrated heterogeneous datasets across multiple published cohorts on all published longitudinal research concerning the course of cognition after TBI. We found that ML captured GII as a predictor of the rate of change across several models in both injury severities in both bivariate and multivariate ML models. The positive correlation between GII and rate of change was consistent with previous reports of the important role of gender equality in heath outcomes 55 . These results are consistent with recent observations showing that trust in relationship are associated with outcomes in advanced clinical conditions, including stroke 56 and dementia 57 . In countries with social constraints on education and development, captured via GII and reflected in economic imbalances and restricted decision making, the macro-level relational environment may impose strains on both family and community relationships, impacting trust. While this may be perceived as far-fetched, results bring attention to the critical role of macro-level social parameters within existing data hierarchies and raise fundamental questions about their relative importance compared with person-level predictors. Greater emphasis on social pathways in future research will be essential to elucidate the mechanisms through which these parameters exert their effects on cognitive recovery. Our correlation matrices results indicate that PROGRESS-Plus parameters characterising study samples, including whether the country of study origin is predominantly English speaking, gender/sex, education, and age are associated with rate of change. However, when these variables were used in ML, in consideration with other parameters (i.e., time effects and country-level structural indicators) their predictive values were not as strong as those of other parameters, except education and age. We have observed implications of the country-level structural indicators expressed by GII in results of our past systematic review 8 , where when we positioned results along the GII continuum, differences between male and females across outcomes and countries of publication started to dissipate. The GII is a composite, time-varying indicator of a country’s gender inequality level, incorporating measures of educational attainment, labour force participation, maternal mortality, adolescent fertility, and parliamentary representation 58 . Therefore, when the index was used in ML models, the predictive effect of cohorts’ education and education SD were diminished, and several of our models featured the GII as a salient predictor of the rate of cognitive change after TBI, in both mild and moderate-severe samples. While it has been previously suggested that cognitive outcomes after TBI differ between people of different gender/sex and education level, our results are the first to highlight that the importance of these parameters are affected by broader social and structural contexts. This has important research, clinical, and policy implications, particularly in ongoing debates regarding the long-term cognitive consequences of milder forms of TBI, and the cumulative contributions of biological sex, gender, education, and injury severity to TBI outcomes. Our ML results also bring attention to the intersectionality reflected in the GII, which may also have implications for those who participated in research included in the dataset, and therefore was captured by ML as a prognostic parameter after TBI. Evidence increasingly shows that the characteristics of research participants in TBI studies are not truly reflective of TBI populations, where racialized, less educated, and more disadvantaged communities are less like to participate in research but are also disproportionately affected by TBI and its adverse consequences 15 , 59 , 60 . Future prognostic research and policy should diversify research samples and strive for systematic inclusion of diverse groups of people in TBI research. Overcoming barriers to research remains a priority 59 , 60 . The purpose of our research was to delineate whether social parameters associated with rate of change of cognition using data from published TBI cohorts, and to provide information on whether the characteristics of longitudinal study samples play a role in prognosis. We applied three ML approaches to build prognostic models concerning rate of cognitive change in both mild and moderate-severe injury severity samples. All three ML approaches converge on the conclusion that social and structural characteristics of published cohorts are prognostically meaningful for cognitive change after TBI. Models for mild TBI strata were more heterogeneous and prone to social influences compared to moderate-severe TBI strata, highlighted in the complexity reflected in the tines hyperparameters of boosting models (GB and XGBoost). Nonetheless, all models indicated that sample compositions are determining prognostic patterns and therefore should be treated with greater care in future research. Results of ML models not only delineated four PROGRESS-Plus characteristics (i.e., country of English language dominance as place of residence, gender/sex, education, and age) of cohorts as important determinants of prognosis, but also exposed critical gaps in the current prognostic evidence base by showing who contributes to longitudinal research and which social variables are measured in cohort studies on prognosis and which are not (i.e., religion/spirituality, socioeconomic status, occupation, social capital), constraining what can be known about the value of these parameters to the course of cognition and prognosis after TBI. Therefore, ML approaches as a methodological tool in research is of great value not only for improving prognostic modelling and meta-research, but also to bring attention to how methodological choices made by researchers about sampling 61 , anchored time scale (i.e., time zero and elapsed time) 62 , and cognitive measures 1 , 53 that actively shaped the current best state of knowledge available to clinicians and policymakers. One limitation of our study is limited reporting of study participants’ social profiles in the dataset of published studies. We developed and implemented rigorous data harmonization processes to maximise the ability to examine all available PROGRESS-Plus characteristics in the dataset prior to application of ML-based prognostic modelling. Despite our data harmonization efforts, the dataset exhibited substantial missingness across several PROGRESS-Plus parameters, including race/ethnicity, socioeconomic status, occupation, religion/spirituality, and social capital. Although data augmentation using synthetic data generation is sometimes employed to mitigate this issue in ML, implementation of this approach was not possible in the present study, as these PROGRESS-Plus parameters were sparsely reported, and the available data was less than 22% for occupation and less than 5% for all other parameters after data harmonization (Table 1 ). At such low levels of representation, synthetic data generation would have relied on insufficient underlying distributions, increasing the risk of amplifying noise, introducing artificial patterns, and reinforcing existing biases. For these reasons, we restricted model training to four parameters with adequate representation: the country of study origin through country of English language dominance, gender/sex, education, and age. Future research should consider this limitation and commit to a standardised reporting of PROGRESS-Plus characteristics to enable equitable and robust ML-based prognostic modelling using published data. Further investigations are also needed to refine the role of PROGRESS-Plus parameters as predictors of outcomes when considering structural level parameters. Our research and results provide a foundation for advancing prediction models that systematically evaluate how study sample representativeness affect TBI prognosis. We employed three ML models, including the RF, the GB, and the XGBoost. Employing SHAP in the present study provided additional insights into model performance and predictive value of each parameter in the presence of others, in both mild and moderate-severe injury severity study samples. The usability of the models in predicting outcomes using published data depends on data quality. We recommend that future researchers and who will apply our process and work with published data dedicate significant time to data extraction, pre-processing and standardization prior to machine learning model input. In addition, users are expected to have at least some machine learning expertise and be familiar with the variables in the dataset, as well as with the principles of modelling. While this may limit direct use by non-technical personnel, it ensures reliable and accurate predictions for research and policy applications. Future enhancements in prognosis research that can be used in machine learning require the systematic collection, reporting, and integration of social parameters in research to enable more equitable clinical and policy decision-making. Methods We conducted and reported our research following the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis – Artificial Intelligence (TRIPOD-AI) guidelines for prediction model development and validation 63 . We shared all research-related files in an open data repository on the Open Science Framework (OSF) 64 , in alliance with the FAIR Guiding Principles 65 . The complete dataset is available upon reasonable request from the corresponding author. Data sources and description The dataset used in this work was developed by our research team using data from our prior systematic reviews 1 , 53 on the course and predictors of cognitive outcomes after TBI. We updated this dataset with new evidence that emerged since the original searches 66 . The flow diagram of the study selection is depicted in Fig. 1 . In brief, the data covers evidence that emerged from five databases: Embase, Medline, Scopus, and Cochrane Central Register for Controlled Trials, and PsycINFO. We searched the databases from inception until April 8, 2024 for peer-reviewed English-language longitudinal studies reporting raw cognitive test scores at two or more time points in adults diagnosed with TBI. Only studies that provided data on the course of cognition over at least two data points after TBI were included. The protocol, PROSPERO registry, and the systematic reviews that provided specifics on study methodology and the dataset are published 15 , 67 , 68 . Our research concerned data on TBI diagnosis, injury severity and cause of injury data, and PROGRESS-Plus parameters of each study sample (country of recruitment, race/language/ethnicity, occupation, gender/sex, religion/spirituality, education, socioeconomic status, social capital, age, and other contextual parameters), and cognitive test scores at baseline and last follow-up assessments 68 . Each cognitive test in a given cohort was treated as a unit of analysis. Dataset preparation and machine learning workflow Figure 1 represents dataset preparation and machine learning workflow, outlining seven steps: (i) data collection; (ii) data extraction; (iii) data harmonization; (iv) data preprocessing; (v) machine learning algorithms; (vi) model evaluation; and (vii) sensitivity analysis, which allowed generation of overall and model-specific explanations. Data collection Study selection followed predefined inclusion and exclusion criteria. Eligible studies included those that reported longitudinal data on standardised cognitive tests and provided sufficient injury-related data of study sample characteristics to support data harmonization and modeling. For more information, please refer to published studies 1 . Data extraction Two independent researchers extracted data using standardized, previously published extraction forms 67 . Extracted variables included demographic characteristics (age, sex/gender, race/ethnicity), social parameters (place of residence, occupation, education, and social capital), and contextual brain injury-related parameters (injury severity, mechanism of injury, comorbidities, and time from injury to baseline assessment). We also extracted data on time from injury to baseline assessment, and calculated the time interval between baseline and last follow-up cognitive assessments. Data was extracted from the main text and supplementary materials of each published study by two independent researchers. Extraction accuracy and reliability were ensured through double extraction, cross-checking, and resolution of discrepancies by consensus. Data harmonization process We considered using ML-based solutions for data harmonization of several variables 69 – 71 . However, we faced significant challenges in finding scientific guidance on the application of standardized ML data harmonization in practice. We found that each harmonization case was a unique process and necessitated a series of context-dependent decisions through iterative team discussions where we evaluated several harmonization options at once (i.e., flexible harmonization 72 , 73 ). We first created a library of terms based on all extracted data (Supplement Table 2 ) that defined variable names, operational definitions, categories, and units to ensure consistency across included cohorts. The core team discussed each variable in the table and evaluated different variables under the same section header, jointly deciding how to incorporate these updates into the harmonized protocol template. We then harmonised data to reconcile different operationalisations of similar concepts (i.e., gender/sex composition of study samples, reported in a variety of data formats, known as syntax) as well as semantics (i.e., intended meanings of words such as young adults, primary schooling, etc.), conceptual schema (i.e., structured and unstructured data extracted from tables and raw text, respectively), and measurement differences (e.g. when the same concept was reported with different measurements, including injury severity, cognitive domain, etc.). Finally, we transformed data to numeric values in order to prepare the data for ML (Supplement Table 3 ). We structured extracted social and contextual elements using the PROGRESS-Plus framework. The categorical parameters of study samples were transformed to continuous variables where feasible: gender/sex composition was transformed to proportion men/males, and education level was transformed to education in years. Additionally, we transformed age-related parameters to continuous age and age standard deviation of study samples. Country of study origin (where recruitment took place) was transformed to country of English language dominance as a binary variable (0, 1). Cognitive outcome scores of standardised tests at baseline and last follow-up and the domains of cognition they reflect were represented as mean and standard deviation, and binary values, respectively. To ensure quality, data harmonization and transformation processes were conducted by two independent researchers. We documented the process and decisions (Supplement Table 3 ) made for future data users in compliance with the FAIR standards 65 . Data preprocessing We conducted data preprocessing. This step included scaling of continuous variables, systematic assessment of outliers, and correction for imbalances across injury severity groups based on visual plots and correlation matrices between predictors and the outcome. When time was reported categorically or as ranges, we used midpoint values. For cohorts reporting multiple follow-up time points 1 , the final available follow-up assessment was used for data harmonization. We calculated time SD for each data point, considering cohort sample size and reported outcome score, and score SDs in order to preserve the variance structure across included cohorts. We then calculated rate of change per month as the outcome for longitudinal harmonisation of data points. For each data point, the rate of change was computed by dividing the difference between outcome values at baseline and last follow-up by the elapsed time between assessments using the following formula: $$\:\text{R}\text{a}\text{t}\text{e}\:\text{o}\text{f}\:\text{c}\text{o}\text{g}\text{n}\text{i}\text{t}\text{i}\text{v}\text{e}\:\text{c}\text{h}\text{a}\text{n}\text{g}\text{e}\:\left(\text{p}\text{e}\text{r}\:\text{m}\text{o}\text{n}\text{t}\text{h}\right)=\frac{{Y}_{FU-}{Y}_{BL}}{T/30}\mathbf{}\mathbf{}$$ where Y FU denotes the cognitive test score at the last follow-up assessment, Y BL denotes the cognitive test score at baseline, and T denotes the mean follow-up interval measured in days. The follow-up interval was converted to months by dividing by 30 to harmonize time scales across studies. Class imbalance We observed that imbalance in our data occurred within injury severity categories. To mitigate the influence of this imbalance on model performance and to preserve clinically meaningful distinctions, we stratified analyses by injury severity rather than applying resampling-based class-imbalance methods. Moderate, moderate–severe, and severe injuries were combined into a single ‘moderate–severe’ category, and models were trained separately for mild and moderate–severe TBI. Missing data We made a priori decision to restrict ML analyses to complete cases, where only data points with complete information on the variables required for ML modelling were included in the final analytic sample 74 – 76 . This decision was made because many variables expressed structurally missing data reporting (Table 2 ) rather than missing at random. Machine learning modeling We applied three supervised ML algorithms 77 to mild and moderate-severe injury severity datasets: random forest (RF) 78 , 79 , gradient boosting (GB) 80 , 81 , and extreme gradient boosting (XGBoost) 82 , to model cognitive change over time. These models were trained to predict both overall cognitive trajectories (standardised rate of change per month) and domain-specific standardised rate of change using harmonised PROGRESS-Plus variables. To ensure comparability across algorithms and mitigate bias introduced by differences in cohorts’ age, time between injury and baseline assessment, and follow-up times, all models were trained on harmonised rates of change per unit of time, and controlled for time between TBI and baseline assessment and time between assessments. We considered time from injury to baseline assessment, time between baseline assessment and last follow-up, and country-level structural indicators as potential modifiers of the relationship between primary predictors and cognitive outcomes. This decision was made in order to capture historical evolution of populations’ social parameters by country, including the GII and HDI. Model evaluation and interpretation We evaluated model performance using mean absolute error (MAE) and root mean square error (RMSE) 83 . We then examined feature importance using feature importance heatmaps and the Shapley Additive Explanations (SHAP) 84 . By observing SHAP values, we evaluated the predictive importance of each feature and how it contributes to the difference between an actual prediction and a mean prediction, analyzing non-linear relationships between PROGRESS-Plus variables and rate of change. Internal validation and sensitivity analysis We performed sensitivity analyses using 5-fold cross-validation as an internal validation approach to evaluate the robustness and generalizability of model estimates. This procedure assessed the stability of the effects of input parameters on cognitive outcomes across injury severity strata. We also conducted sensitivity analyses to assess the potential effects of PROGRESS-Plus parameters on specific cognitive domains that provided sufficient data and had relatively even distribution between mild and moderate-severe TBI cohorts. We hypothesized a priori that PROGRESS-Plus parameters would exert a more pronounced effect on domain-specific cognitive outcomes than on overall cognition, stronger for mild than moderate-severe TBI, given previously observed variability in rates of change by injury severity within the cohorts 1 . This allowed us to verify the relative contributions of the important PROGRESS-Plus parameters under a more constrained, cognitive domain-specific dataset. We re-evaluated feature importance using feature importance heatmaps and the SHAP, estimating each predictor’s contribution to standardised rate of cognitive change in cognitive domain-specific datasets for executive function and learning and memory domains of cognition. We also investigated the model performance by removing features with high correlation coefficients to evaluate the impact of multicollinearity on the model’s performance. By selectively excluding highly correlated features (education and age SD), we aimed to improve the robustness of the model. All analyses, including generation of figures, were performed using Python (version 3) 85,86 . Figures were created using Python libraries (matplotlib (v3.x), seaborn (v0.x), and wordcloud (v1.x)). Declarations The study protocol was approved by the ethics committees at the clinical (Toronto Rehabilitation Institute–University Health Network) and academic (University of Toronto) institutions. All methods were carried out in accordance with the relevant guidelines and regulations. Informed consent: This research utilised published data with no access to personal information. Competing Interests Statement The authors declare no competing interests. Ethical approval and informed consent Informed consent: This research utilised published data with no access to personal information. Funding CIHR Brain Health and Reduction of Risk for Age-related Cognitive Impairment SGD #202306BK5-510306-BKS-ADHD-220229, and Canada Research Chair in Neurological Disorders and Brain Health, CRC-2021-00074. Author Contribution TM and TTS conceived the project. TM supervised US and TTS data extraction and harmonization. TM developed templates for data library encoding and transformation processing. JX and US performed the data analyses with methodological supervision from TM. US created visual display of data. TM wrote the first draft of the manuscript with input from US. All authors contributed to data interpretation and editing of the paper. Acknowledgement The authors thank Professor Michael Escobar for contribution to the technical discussion on mathematical evolution of the concept; trainees of the BRIDGE Lab (bridgelab.ca), Ashlee Kim, Chuxi Pan, and Mursal Jahed, for their support with title and abstract screening and data extraction; and Information Specialists from Library Services at the University Health Network (Jessica Babineau, Emilia Main, and Cynthia Chui) for help in retrieving data for systematic reviews that formed the basis for the data used in this research.This study was funded by the Canadian Institutes of Health Research (CIHR) Brain Health and Reduction of Risk for Age-related Cognitive Impairment SGD #202306BK5-510306-BKS-ADHD-220229, and in part by Canada Research Chair in Neurological Disorders and Brain Health, CRC-2021-00074. This research program was developed with endorsement by patient and public organizations, who reviewed the research program prior to its submission to the funding agency (Canadian Institute of Health Research). The critical need to address PROGRESS-Plus parameters in brain injury research was also endorsed by team members of the SGD #202306BK5-510306-BKS-ADHD-220229. The content is solely the authors' responsibility and does not necessarily represent the official views of the CIHR. Data Availability The data encoding and harmonisation templates used in this research are available from the public [data portal](https:/osf.io/jku23/overview) . The full dataset creation plan and underlying analytic code are available from the corresponding author upon request. References Mollayeva, T., Mollayeva, S., Pacheco, N., D’Souza, A. & Colantonio, A. The course and prognostic factors of cognitive outcomes after traumatic brain injury: A systematic review and meta-analysis. Neurosci. Biobehav Rev. 99 , 198–250 (2019). Tritt, A. et al. Data-driven distillation and precision prognosis in traumatic brain injury with interpretable machine learning. Sci. Rep. 13 , 21200 (2023). Bark, D. et al. Refining outcome prediction after traumatic brain injury with machine learning algorithms. Sci. Rep. 14 , 8036 (2024). Oldenburg, C., Lundin, A., Edman, G., Nygren-de Boussard, C. & Bartfai, A. Cognitive reserve and persistent post-concussion symptoms—A prospective mild traumatic brain injury (mTBI) cohort study. Brain Inj . 30 , 146–155 (2016). Theadom, A. et al. Social cognition four years after mild-TBI: An age-matched prospective longitudinal cohort study. Neuropsychology 33 , 560–567 (2019). HANAFY, S. et al. Comorbidity in traumatic brain injury and functional outcomes: a systematic review. Eur J. Phys. Rehabil Med 57 , (2021). Büttner, F., Terry, D. P. & Iverson, G. L. Using a Likelihood Heuristic to Summarize Conflicting Literature on Predictors of Clinical Outcome Following Sport-Related Concussion. Clin. J. Sport Med. 31 , e476–e483 (2021). Mollayeva, T., Mollayeva, S., Pacheco, N. & Colantonio, A. Systematic Review of Sex and Gender Effects in Traumatic Brain Injury: Equity in Clinical and Functional Outcomes. Front Neurol 12 , (2021). McIntosh, S. J., Vergeer, M. H., Galarneau, J. M., Eliason, P. H. & Debert, C. T. Factors Associated With Persisting Symptoms After Concussion in Adults With Mild TBI. JAMA Netw. Open. 8 , e2516619 (2025). Mollayeva, T. et al. Decoding health status transitions of over 200 000 patients with traumatic brain injury from preceding injury to the injury event. Sci. Rep. 12 , 5584 (2022). Mollayeva, T., Tran, A., Chan, V., Colantonio, A. & Escobar, M. D. Sex-specific analysis of traumatic brain injury events: applying computational and data visualization techniques to inform prevention and management. BMC Med. Res. Methodol. 22 , 30 (2022). Teterina, A. et al. Gender versus sex in predicting outcomes of traumatic brain injury: a cohort study utilizing large administrative databases. Sci. Rep. 13 , 18453 (2023). Yaniv, D. Trust the Process: A New Scientific Outlook on Psychodramatic Spontaneity Training. Front Psychol 9 , (2018). Hanazaki, N. Local and traditional knowledge systems, resistance, and socioenvironmental justice. J. Ethnobiol. Ethnomed. 20 , 5 (2024). Kim, H., Jahed, M., Sant’Ana, T. T. & Mollayeva, T. (eds) (Ashlee), Cognitive Outcomes Among Adults Following Traumatic Brain Injury: A Systematic Review Using the PROGRESS-plus Framework. Arch Phys Med Rehabil 106, e120 (2025). Bleiberg, J. et al. Duration of Cognitive Impairment After Sports Concussion. Neurosurgery 54 , 1073–1080 (2004). Chen, C. C., Wei, S. T., Tsaia, S. C., Chen, X. X. & Cho, D. Y. Cerebrolysin enhances cognitive recovery of mild traumatic brain injury patients: double-blind, placebo-controlled, randomized study. Br. J. Neurosurg. 27 , 803–807 (2013). Covassin, T., Stearne, D. & Elbin, R. Concussion History and Postconcussion Neurocognitive Performance and Symptoms in Collegiate Athletes. J. Athl Train. 43 , 119–124 (2008). Dikmen, S., Machamer, J. & Temkin, N. Mild Traumatic Brain Injury: Longitudinal Study of Cognition, Functional Status, and Post-Traumatic Symptoms. J. Neurotrauma . 34 , 1524–1530 (2017). Farbota, K. D. et al. Longitudinal diffusion tensor imaging and neuropsychological correlates in traumatic brain injury patients. Front Hum. Neurosci 6 , (2012). Field, M., Collins, M. W., Lovell, M. R. & Maroon, J. Does age play a role in recovery from sports-related concussion? A comparison of high school and collegiate athletes. J. Pediatr. 142 , 546–553 (2003). Hicks, A. J. et al. Does cognitive decline occur decades after moderate to severe traumatic brain injury? A prospective controlled study. Neuropsychol. Rehabil . 32 , 1530–1549 (2022). Kersel, D. A., Marsh, N. V., Havill, J. H. & Sleigh, J. W. Neuropsychological functioning during the year following severe traumatic brain injury. Brain Inj . 15 , 283–296 (2001). Kontos, A. P. et al. The effects of combat-related mild traumatic brain injury (mTBI). J. Trauma. Acute Care Surg. 79 , S146–S151 (2015). Kwok, F. Y., Lee, T. M. C., Leung, C. H. S. & Poon, W. S. Changes of cognitive functioning following mild traumatic brain injury over a 3-month period. Brain Inj . 22 , 740–751 (2008). Lee, H. et al. Comparing effects of methylphenidate, sertraline and placebo on neuropsychiatric sequelae in patients with traumatic brain injury. Hum. Psychopharmacology: Clin. Experimental . 20 , 97–104 (2005). Lu, J., Roe, C., Sigurdardottir, S., Andelic, N. & Forslund, M. Trajectory of Functional Independent Measurements during First Five Years after Moderate and Severe Traumatic Brain Injury. J. Neurotrauma . 35 , 1596–1603 (2018). Macciocchi, S. N., Seel, R. T. & Thompson, N. The Impact of Mild Traumatic Brain Injury on Cognitive Functioning Following Co-occurring Spinal Cord Injury. Arch. Clin. Neuropsychol. 28 , 684–691 (2013). Malec, J. F. et al. Longitudinal Effects of Medical Comorbidities on Functional Outcome and Life Satisfaction After Traumatic Brain Injury: An Individual Growth Curve Analysis of NIDILRR Traumatic Brain Injury Model System Data. J. Head Trauma. Rehabilitation . 34 , E24–E35 (2019). Mandleberg, I. A. Cognitive recovery after severe head injury. 3. WAIS verbal and performance IQs as a function of post-traumatic amnesia duration and time from injury. J. Neurol. Neurosurg. Psychiatry . 39 , 1001–1007 (1976). Paré, N., Rabin, L. A., Fogel, J. & Pépin, M. Mild traumatic brain injury and its sequelae: Characterisation of divided attention deficits. Neuropsychol. Rehabil . 19 , 110–137 (2009). Robertson, K. & Schmitter-Edgecombe, M. Self-awareness and traumatic brain injury outcome. Brain Inj . 29 , 848–858 (2015). Sandhaug, M., Andelic, N., Langhammer, B. & Mygland, A. Functional level during the first 2 years after moderate and severe traumatic brain injury. Brain Inj . 29 , 1431–1438 (2015). Schmitter-Edgecombe, M. & Robertson, K. Recovery of visual search following moderate to severe traumatic brain injury. J. Clin. Exp. Neuropsychol. 37 , 162–177 (2015). SNOW, P., PONSFORD, J. & DOUGLAS, J. & Conversational discourse abilities following severe traumatic brain injury: a follow up study. Brain Inj . 12 , 911–935 (1998). Vanderploeg, R. D., Donnell, A. J., Belanger, H. G. & Curtiss, G. Consolidation deficits in traumatic brain injury: The core and residual verbal memory defect. J. Clin. Exp. Neuropsychol. 36 , 58–73 (2014). Wang, S. et al. Umbilical cord mesenchymal stem cell transplantation significantly improves neurological function in patients with sequelae of traumatic brain injury. Brain Res. 1532 , 76–84 (2013). Zaninotto, A. L. et al. Visuospatial memory improvement in patients with diffuse axonal injury (DAI): a 1-year follow-up study. Acta Neuropsychiatr . 29 , 35–42 (2017). Prigatano, G. P. et al. Neuropsychological rehabilitation after closed head injury in young adults. J. Neurol. Neurosurg. Psychiatry . 47 , 505–513 (1984). Sours, C. et al. Associations between interhemispheric functional connectivity and the Automated Neuropsychological Assessment Metrics (ANAM) in civilian mild TBI. Brain Imaging Behav. 9 , 190–203 (2015). Stenberg, J. et al. Change in self-reported cognitive symptoms after mild traumatic brain injury is associated with changes in emotional and somatic symptoms and not changes in cognitive performance. Neuropsychology 34 , 560–568 (2020). Wylie, G. R. et al. Cognitive Improvement after Mild Traumatic Brain Injury Measured with Functional Neuroimaging during the Acute Period. PLoS One . 10 , e0126110 (2015). Powell, T. J., Collin, C. & Sutton, K. A follow-up study of patients hospitalized after minor head injury. Disabil. Rehabil . 18 , 231–237 (1996). Barker-Collo, S. et al. Neuropsychological outcome and its correlates in the first year after adult mild traumatic brain injury: A population-based New Zealand study. Brain Inj . 29 , 1604–1616 (2015). McCREA, M. et al. Standard regression-based methods for measuring recovery after sport-related concussion. J. Int. Neuropsychol. Soc. 11 , 58–69 (2005). Rajkomar, A., Hardt, M., Howell, M. D., Corrado, G. & Chin, M. H. Ensuring Fairness in Machine Learning to Advance Health Equity. Ann. Intern. Med. 169 , 866–872 (2018). Murman, D. The Impact of Age on Cognition. Semin Hear. 36 , 111–121 (2015). Ayton, A., Spitz, G., Hicks, A. J. & Ponsford, J. Ageing with Traumatic Brain Injury: Long-Term Cognition and Wellbeing. J. Neurotrauma . https://doi.org/10.1089/neu.2024.0524 (2025). Kirkwood, T. B. L. & Austad, S. N. Why do we age? Nature 408 , 233–238 (2000). Kirkwood, T. B. L. Evolution of ageing. Nature 270 , 301–304 (1977). Chan, V., Mollayeva, T., Ottenbacher, K. J. & Colantonio, A. Sex-Specific Predictors of Inpatient Rehabilitation Outcomes After Traumatic Brain Injury. Arch. Phys. Med. Rehabil . 97 , 772–780 (2016). Mollayeva, T., Hurst, M., Escobar, M. & Colantonio, A. Sex-specific incident dementia in patients with central nervous system trauma. Alzheimer’s Dementia: Diagnosis Assess. Disease Monit. 11 , 355–367 (2019). D’Souza, A. et al. Measuring Change Over Time: A Systematic Review of Evaluative Measures of Cognitive Functioning in Traumatic Brain Injury. Front Neurol 10 , (2019). Bryant, A. M. et al. Profiles of Cognitive Functioning at 6 Months After Traumatic Brain Injury Among Patients in Level I Trauma Centers. JAMA Netw. Open. 6 , e2349118 (2023). Armstrong, K. & Ritchie, C. Research Participation in Marginalized Communities — Overcoming Barriers. N. Engl. J. Med. 386 , 203–205 (2022). Marshall, I. J. et al. The effects of socioeconomic status on stroke risk and outcomes. Lancet Neurol. 14 , 1206–1218 (2015). Ornish, D. et al. Effects of intensive lifestyle changes on the progression of mild cognitive impairment or early dementia due to Alzheimer’s disease: a randomized, controlled clinical trial. Alzheimers Res. Ther. 16 , 122 (2024). Gender Inequality Index (GII). United Nations Development Program: Human Development Reports https://hdr.undp.org/data-center/thematic-composite-indices/gender-inequality-index#/indicies/GII Cookson, R., Robson, M., Skarda, I. & Doran, T. Equity-informative methods of health services research. J. Health Organ. Manag . 35 , 665–681 (2021). Castillo, E. G. & Harris, C. Directing Research Toward Health Equity: a Health Equity Research Impact Assessment. J. Gen. Intern. Med. 36 , 2803–2808 (2021). Graziano, F., Valsecchi, M. G. & Rebora, P. Sampling strategies to evaluate the prognostic value of a new biomarker on a time-to-event end-point. BMC Med. Res. Methodol. 21 , 93 (2021). Kent, P., Cancelliere, C., Boyle, E., Cassidy, J. D. & Kongsted A. A conceptual framework for prognostic research. BMC Med. Res. Methodol. 20 , 172 (2020). Collins, G. S. et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385 , e078378 (2024). Tatyana Mollayeva, D., Gi, E. M. E. E. & F.-T. D. G. S. H. B. L. M. P. T. T. S. D. C. Data Templates in A PROGRESS-Driven Approach to Cognitive Outcomes after Traumatic Brain Injury: Advancing Equity, Diversity, and Inclusion through Knowledge Synthesis and Mobilization. OSF (2026). Wilkinson, M. D. et al. The FAIR Guiding Principles for scientific data management and stewardship. Sci. Data . 3 , 160018 (2016). Sant’Ana, T. T. et al. A PROGRESS-driven approach to cognitive outcomes after traumatic brain injury: A study protocol for advancing equity, diversity, and inclusion through knowledge synthesis and mobilization. PLoS One . 19 , e0307418 (2024). Mollayeva, T., Pacheco, N., D’Souza, A. & Colantonio, A. The course and prognostic factors of cognitive status after central nervous system trauma: a systematic review protocol. BMJ Open. 7 , e017165 (2017). Mollayeva, T. et al. C. (). A systematic review of brain health outcomes after traumatic brain injury through a PROGRESS-Plus lens. PROSPERO (2024). Voß, H. et al. HarmonizR enables data harmonization across independent proteomic datasets with appropriate handling of missing values. Nat. Commun. 13 , 3523 (2022). Adhikari, K. et al. Data Harmonization and Data Pooling from Cohort Studies: A Practical Approach for Data Management. Int J. Popul. Data Sci 6 , (2021). Gill, I. S. et al. The DataHarmonizer: a tool for faster data harmonization, validation, aggregation and analysis of pathogen genomics contextual information. Microb Genom 9 , (2023). Fortier, I., Doiron, D., Burton, P. & Raina, P. Invited Commentary: Consolidating Data Harmonization–How to Obtain Quality and Applicability? Am. J. Epidemiol. 174 , 261–264 (2011). Boyden, J. & Walnicki, D. Leveraging the Power of Longitudinal Data: Insights on Data Harmonisation and Linkage from Young Lives . (2021). Löw, N., Hesser, J. & Blessing, M. Multiple retrieval case-based reasoning for incomplete datasets. J. Biomed. Inf. 92 , 103127 (2019). Mainzer, R. M. et al. Gaps in the usage and reporting of multiple imputation for incomplete data: findings from a scoping review of observational studies addressing causal questions. BMC Med. Res. Methodol. 24 , 193 (2024). Lee, K. J., Carlin, J. B., Simpson, J. A. & Moreno-Betancur, M. Assumptions and analysis planning in studies with missing data in multiple variables: moving beyond the MCAR/MAR/MNAR classification. Int. J. Epidemiol. 52 , 1268–1275 (2023). Lundberg, S. M. et al. From local explanations to global understanding with explainable AI for trees. Nat. Mach. Intell. 2 , 56–67 (2020). Hu, J. & Szymczak, S. A review on longitudinal data analysis with random forest. Brief Bioinform 24 , (2023). Wallace, M. L. et al. Use and misuse of random forest variable importance metrics in medicine: demonstrations through incident stroke prediction. BMC Med. Res. Methodol. 23 , 144 (2023). Lee, Y. W., Choi, J. W. & Shin, E. H. Machine learning model for predicting malaria using clinical information. Comput. Biol. Med. 129 , 104151 (2021). Austin, P. C., Harrell, F. E., Lee, D. S. & Steyerberg, E. W. Empirical analyses and simulations showed that different machine and statistical learning methods had differing performance for predicting blood pressure. Sci. Rep. 12 , 9312 (2022). Lee, H. S., Kim, J. H., Son, J., Park, H. & Choi, J. Machine learning models for predicting early hemorrhage progression in traumatic brain injury. Sci. Rep. 14 , 11690 (2024). Robeson, S. M. & Willmott, C. J. Decomposition of the mean absolute error (MAE) into systematic and unsystematic components. PLoS One . 18 , e0279774 (2023). Jiang, S., Liang, Y., Shi, S., Wu, C. & Shi, Z. Improving predictions and understanding of primary and ultimate biodegradation rates with machine learning models. Sci. Total Environ. 904 , 166623 (2023). Pedregosa, F. Machine learning in python - extra trees regressor. Scikit Learn (2011). https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.ExtraTreesRegressor.html Pedregosa, F. Machine learning in python - gradient boosting regressor. Scikit Learn (2011). https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.GradientBoostingRegressor.html Tables Tables are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1PROGRESS.docx Table2Modelperformance.docx Table3Modelhyperparameters.docx Table45CVModelperformance.docx Supplementarymaterial.docx Cite Share Download PDF Status: Published Journal Publication published 30 Mar, 2026 Read the published version in Scientific Reports → 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-8585532","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":578326003,"identity":"7332b9f9-2634-4abf-9589-1b9ab1a598d0","order_by":0,"name":"Jingwen Xu","email":"","orcid":"","institution":"University Health Network","correspondingAuthor":false,"prefix":"","firstName":"Jingwen","middleName":"","lastName":"Xu","suffix":""},{"id":578326004,"identity":"845971b9-ff7d-4a86-a3c0-d8b03cb49294","order_by":1,"name":"Urooba Shaikh","email":"","orcid":"","institution":"University Health Network","correspondingAuthor":false,"prefix":"","firstName":"Urooba","middleName":"","lastName":"Shaikh","suffix":""},{"id":578326005,"identity":"321e4c20-e322-4a7d-8468-6d98b6f67e87","order_by":2,"name":"Thaisa Tylinski Sant’Ana","email":"","orcid":"","institution":"University Health Network","correspondingAuthor":false,"prefix":"","firstName":"Thaisa","middleName":"Tylinski","lastName":"Sant’Ana","suffix":""},{"id":578326006,"identity":"a01422d6-ef46-4dc4-92b2-90138ac70abf","order_by":3,"name":"Tatyana Mollayeva","email":"data:image/png;base64,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","orcid":"","institution":"University Health Network","correspondingAuthor":true,"prefix":"","firstName":"Tatyana","middleName":"","lastName":"Mollayeva","suffix":""}],"badges":[],"createdAt":"2026-01-12 22:23:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8585532/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8585532/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-44818-5","type":"published","date":"2026-03-30T15:58:43+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":101269055,"identity":"6a52e4c3-5209-4d5d-97f5-36e8bbdd09f5","added_by":"auto","created_at":"2026-01-28 01:27:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":10509277,"visible":true,"origin":"","legend":"\u003cp\u003eStudy design and analytical workflow. (Attached as TIFF file)\u003c/p\u003e\n\u003cp\u003eFigure 1 legend. Schematic representation of the analytical pipeline, illustrating the flow from data acquisition to model interpretation. The seven steps include data collection, extraction, harmonization, preprocessing, machine-learning model development, model evaluation, and sensitivity analysis, with intermediate and final results displayed at each stage to support overall and model-specific explanations.\u003c/p\u003e","description":"","filename":"Fig1MLworkflow.png","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/57c86aff7a338829737732fb.png"},{"id":101296873,"identity":"a5ca42f3-8d0e-4343-b6d7-5835cfcc4e41","added_by":"auto","created_at":"2026-01-28 09:22:21","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":616478,"visible":true,"origin":"","legend":"\u003cp\u003eHeatmaps of the correlation coefficient matrix, by injury severity\u003c/p\u003e\n\u003cp\u003elegend. Heatmaps of the correlation coefficient matrix, for mild (a) and moderate-severe (b) TBI for each of the parameters’ association with the outcome, rate of change (bivariate associations). Red signifies a positive correlation, while blue represents a negative correlation. The intensity of the color reflects the magnitude of the correlation coefficient, with more vibrant shades indicating stronger correlations. Specifically, shades tending towards red represent coefficients approaching 1, while those leaning towards blue represent coefficients approaching -1.\u003c/p\u003e","description":"","filename":"Fig2Mainheatmaps.png","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/a0a717d01155356e56fcd9ee.png"},{"id":101269051,"identity":"0a7fe79f-2426-4a12-821c-4b6eceda112f","added_by":"auto","created_at":"2026-01-28 01:27:07","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":782636,"visible":true,"origin":"","legend":"\u003cp\u003eFeature contributions across machine-learning models.\u003c/p\u003e\n\u003cp\u003elegend. Ranking of feature contributions for prediction of the rate of change in cognitive outcomes in (a) mild and (b) moderate-to-severe traumatic brain injury (TBI) cohorts.\u003c/p\u003e","description":"","filename":"Fig3Mainfeatureimportance.png","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/77cf0b2e115b9c9e035dbf50.png"},{"id":101269052,"identity":"da55d4c9-2a42-472b-997e-2ec61b3091a6","added_by":"auto","created_at":"2026-01-28 01:27:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":726577,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP summary of feature impact on predictive models’ output.\u003c/p\u003e\n\u003cp\u003elegend. SHAP summary plots illustrating the influence of input variables on model predictions across models for (a) mild and (b) moderate-severe TBI cohorts, respectively. Each point represents a SHAP value for each input variable, indicating the magnitude and direction of a feature’s contribution to the predicted outcome, summarizing model-specific predictive patterns for each TBI severity group. Features are ordered according to their relative contribution to model predictions.\u003c/p\u003e","description":"","filename":"Fig4MainSHAPplots.png","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/e7a42023edebdff3ae784b52.png"},{"id":106344864,"identity":"10c5d077-2f86-4fe6-9452-178da90eb9a2","added_by":"auto","created_at":"2026-04-07 16:17:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":20290266,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/0da6f2fb-ef82-4f54-bc69-1e7eb2a972b2.pdf"},{"id":101269054,"identity":"3519e3a7-93ba-4a76-bbd4-6dfab2dcc8ae","added_by":"auto","created_at":"2026-01-28 01:27:07","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":15962,"visible":true,"origin":"","legend":"","description":"","filename":"Table1PROGRESS.docx","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/3d628e6b61898ba45f483eb8.docx"},{"id":101269049,"identity":"f033c968-e8fc-4c06-9796-cea3da81d3c0","added_by":"auto","created_at":"2026-01-28 01:27:07","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":15141,"visible":true,"origin":"","legend":"","description":"","filename":"Table2Modelperformance.docx","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/173601d976616e93cea6239e.docx"},{"id":101298096,"identity":"5e989d68-fc78-45b3-a396-21cbf02a07e4","added_by":"auto","created_at":"2026-01-28 09:30:18","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":13838,"visible":true,"origin":"","legend":"","description":"","filename":"Table3Modelhyperparameters.docx","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/41acb09d8140e1a6dbdb9f25.docx"},{"id":101269057,"identity":"8502333a-cf06-4e74-a5f2-7d3fa9cb7994","added_by":"auto","created_at":"2026-01-28 01:27:07","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":12620,"visible":true,"origin":"","legend":"","description":"","filename":"Table45CVModelperformance.docx","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/a3ecbc1492c07a2522b70493.docx"},{"id":101269056,"identity":"bc1c386a-e7a4-4de8-bcf1-88f760d75f61","added_by":"auto","created_at":"2026-01-28 01:27:07","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":14120575,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-8585532/v1/0b2a33cd7609dc91cc5581dd.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Explainable machine learning of PROGRESS-Plus social factors predicts cognitive trajectories after traumatic brain injury","fulltext":[{"header":"Introduction","content":"\u003cp\u003eResearch on the course and prognosis of cognitive function following traumatic brain injury (TBI), which refers to structural and/or physiological disruption of brain function due to an external force\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e, is rapidly developing but remains characterized by extensive inter-patient and intra-patient heterogeneity. Adverse outcomes are attributed to greater injury severity and advanced age, highlighting the challenges in care and poor prognosis for these patients\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Recently, the view of TBI of mild severity is progressing through evidence towards a condition with long-term consequences for cognition\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e rather than a singular event. Furthermore, it has been proposed that TBI should not be seen as a random event and accident, but rather as a health status transition on a continuum, where the magnitude of social adversities and comorbidities increase the risk of injury severity, its adverse course, and expedited\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan additionalcitationids=\"CR7 CR8\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e cognitive decline, across injury severities\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. With this in mind, recent research has started to increasingly pose social equity hypotheses and investigate the role of social parameters in recovery course and long-term outcomes. However, the challenge remains that social parameters are intertwined with injury-related and environmental risks, age-related parameters, and broader social adversities, forming a vastly complex network of associations\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Capturing and disentangling these associations on a time continuum presents a substantial challenge for traditional hypothesis-driven research, which typically assumes linear relationships and examines a limited number of variables\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. As a result, important sources of variability in TBI prognosis remain insufficiently characterised in scientific research, creating a knowledge gap in research, policy, and practice\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eEmerging advances in computational and machine learning (ML) approaches offer a timely opportunity to address this gap by enabling more systematic evaluation of variable importance and the complex interplay of social and demographic parameters in shaping cognitive trajectories after TBI. One method to increase confidence in the prognosis of cognitive course after TBI is to apply computational and ML approaches to longitudinal data from published TBI cohorts. These approaches enable the evaluation of variable importance and the complex interplay of social and demographic parameters shaping cognitive outcomes after TBI. Here, we present a new analysis of data from 30 published longitudinal cohort studies characterising change in cognitive test performance with 2,364 participants with TBI (72% male, mean age 32 years; 55% mild, 45% moderate/severe TBI)\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. We leveraged advancements in computational and ML approaches, with the aim to: (1) develop and validate predictive models of longitudinal cognitive change after TBI using harmonised social, demographic, and injury-related data from published cohort studies; (2) compare the predictive performance of multiple ML approaches (random forest, gradient boosting, and extreme gradient boosting) in modelling standardised rates of cognitive change following TBI, stratified by injury severity, and (3) identify and interpret the relative importance of social and demographic parameters in predicting cognitive trajectories after TBI. Our overarching aim was to determine the prognostic significance of social risk stratification in TBI by leveraging an interpretable ML framework. Through this approach, we sought to move beyond traditional injury-related predictors and elucidate how social profiles of study samples shape longitudinal cognitive course after TBI. This approach to characterizing heterogeneity in cognitive outcomes will equip researchers and policymakers with actionable evidence for understanding and interpreting heterogeneity in cognitive outcomes, and informing more equitable, data-driven prediction.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eWe included 30 published studies\u003csup\u003e16–45\u0026nbsp;\u003c/sup\u003eand 43 cohorts within them (i.e., Mandleberg et al. provided eight cohorts\u003csup\u003e30\u003c/sup\u003e, Dikmen et al. provided three cohorts\u003csup\u003e\u0026nbsp;19\u003c/sup\u003e, and Kontos et al.\u003csup\u003e24\u003c/sup\u003e, Macciochi et al.\u003csup\u003e28\u003c/sup\u003e, and Covassin et al.\u003csup\u003e18\u003c/sup\u003e provided two cohorts, each), which provided a total of 218 cognitive outcome data points (see dataset). The analysis included 218 data points collected on at least two time points after injury comprising 2,364 adults (72% male) with TBI. \u0026nbsp;Of these, 57% corresponded to data of participants with mild TBI, with the remaining data points representing moderate (4.6%), moderate-severe TBI (24%) and severe TBI (16%). To mitigate the influence of imbalance in our data in injury severity categories and to preserve clinically meaningful distinctions, moderate, moderate–severe, and severe injuries were combined into a single ‘moderate–severe’ category, and models were trained separately for mild and moderate–severe TBI.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePublished data spanned five decades, with the majority (87%) published after the year 2000. All studies were conducted in high- and middle-income countries. Most cohorts originated from the USA (53%), followed by New Zealand and UK (9.6% each), China (6.9%), Norway (5.0%), South Korea (4.6%), Brazil (4.1%), and Canada and Australia (3.7%, each) (Table 1). To strengthen the use of country of study origin as a proxy for place of residence, we used country of English language dominance instead of country of study origin in the modelling process, categorizing it under place of residence. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInsert Table 1 here\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData extraction process\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAfter combining moderate and severe injury severity with moderate-severe data points, the final dataset resulted in 121 data points for mild, and 97 data points for moderate-severe TBI (Supplement Table 1).\u003c/p\u003e\n\u003cp\u003ePublished data spanned five decades, with the majority (87%) published after the year 2000. Data points represented nine countries, all of high- and middle-income, across global regions. Most cohorts originated from the USA (53%), followed by New Zealand and UK (9.6% each), China (6.9%), Norway (5.0%), South Korea (4.6%), Brazil (4.1%), and Canada and Australia (3.7%, each) (Table 1). To strengthen the use of country of study origin as a proxy for place of residence, we used country of English language dominance instead of country of study origin in the modelling process, categorizing it under place of residence. \u0026nbsp;Reporting of study sample characteristics varied across cohorts, regardless of country of study origin. Gender/sex composition of study samples was reported in all cohorts, and after harmonisation proportion of males in samples ranged from 0.44\u003csup\u003e42\u003c/sup\u003e to 1\u003csup\u003e16,24\u003c/sup\u003e (Table 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCohorts’ education was reported using a variety of formats. After harmonisation to years of formal education, mean education levels ranged from 7.1\u003csup\u003e17,37\u003c/sup\u003e to 14.7\u003csup\u003e42\u003c/sup\u003e years across cohorts; 11\u003csup\u003e25,26\u003c/sup\u003e to 14.7\u003csup\u003e42\u003c/sup\u003e years for mild, and 7.1\u003csup\u003e17,37\u003c/sup\u003e to 13.17\u003csup\u003e20\u003c/sup\u003e years for moderate-severe TBI cohorts. Education SDs ranged from 1.11\u003csup\u003e35\u003c/sup\u003e to 4\u003csup\u003e\u0026nbsp;17,37\u003c/sup\u003e years, 1.5\u003csup\u003e28\u003c/sup\u003e to 4\u003csup\u003e\u0026nbsp;17,37\u003c/sup\u003e years for mild, and 1.11\u003csup\u003e35\u003c/sup\u003e to 3.3\u003csup\u003e39\u003c/sup\u003e years for moderate-severe TBI cohorts.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOther PROGRESS-Plus variables (i.e., race/ethnicity, occupation, social capital, and socioeconomic status) were represented by a limited number of data points and were not used in data analysis (Supplement Table 1).\u003c/p\u003e\n\u003cp\u003eOf the Plus parameters, age and age SD were available for all data points. The pooled mean age across studies ranged from 18.8\u003csup\u003e16\u003c/sup\u003e to 43.68\u003csup\u003e40\u003c/sup\u003e years, 18.8\u003csup\u003e\u0026nbsp;16\u003c/sup\u003e to 43.68\u003csup\u003e40\u003c/sup\u003e for mild TBI and 25.2\u003csup\u003e36\u003c/sup\u003e to 41\u003csup\u003e33\u003c/sup\u003e years for moderate-severe TBI (please see dataset for specifics and data by injury severity). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe baseline assessment time (time zero) ranged from ~12 hours\u003csup\u003e\u0026nbsp;16\u003c/sup\u003e to ~10 years post-injury\u003csup\u003e22\u003c/sup\u003e and final follow-up assessments ranged from five days\u003csup\u003e18\u003c/sup\u003e to ~20 years\u003csup\u003e22\u003c/sup\u003e post-TBI (Supplement table 1), with the mean number of days from injury to baseline assessment being 11.2 for mild and 348 for moderate-severe TBI cohorts.\u003c/p\u003e\n\u003cp\u003eFor the purpose of observing rate of change per month as the outcome, we then converted time-related variables to the common metric of months by dividing by 30.4, the average number of days in a month (Supplement Table 1). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe distribution of data points across cognitive test domains, overall and by injury severity, was as follows, from highest to lowest: executive function (118; 55% mild), learning and memory (105, 49% mild), perceptual-motor (87, 64% mild), information processing speed (85, 72% mild), complex attention (84, 60% mild), language (27, 41% mild), and social cognition (6, 100% mild). We refer the reader to Supplement Table 1 for data used in this synthesis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe cognitive outcome data for ML was a nominal variable across domains of cognition, standardised as rate of change per month (see Methods section).The rate of change was also calculated for domains of cognition with sufficient data: executive function and learning and memory.\u003c/p\u003e\n\u003cp\u003eThe outcome data points were collected at baseline and last follow-up, and formed the foundation for the predictive modelling of cognitive trajectories presented in this study (Supplement Table 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFrom baseline to last follow-up assessment, the median standardised rate of change per month was 0.233 for mild, and 0.112 for moderate-severe TBI. We refer the reader to the figures for results of the standardised rate of change for mild and moderate-severe TBI for each of the PROGRESS-Plus parameters with available data: the country of study origin, gender/sex and education overall, and by domains of cognition (Supplement Figure 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eData preprocessing\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe created heat maps with correlation coefficient matrices for mild and moderate-severe TBI to identify associations among PROGRESS-Plus variables (Figure 2 and Supplement Figure 2 for country-specific associations).\u0026nbsp;We treated all PROGRESS-Plus variables with complete data coverage as primary predictors and included them in the analyses irrespective of the statistical significance of their association with the rate of change.\u0026nbsp;While country of study origin was available for each cohort, in order to use it efficiently (i.e., considering historical evolution) we used Gender Inequality Index (GII) and Human Development Index (HDI) as contextual measures and effect modifiers in the modelling process. These country-level structural indicators contributed to measurable heterogeneity across mild and moderate-severe TBI cohorts in several key predictors (Figure 2). Assigned HDI and GII values ranged from 0.689\u003csup\u003e25\u003c/sup\u003e to 0.928\u003csup\u003e24,32,34,42\u003c/sup\u003e and 0.009\u003csup\u003e41\u003c/sup\u003e to 0.269\u003csup\u003e38\u003c/sup\u003e, respectively. Temporal trends indicated a gradual change in values in cohorts coming from the same country (refer to dataset).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAlong with PROGRESS-Plus parameters, we also evaluated whether time-related variables (i.e., time from injury to baseline assessment and time between baseline and last follow up assessment) were associated with the rate of change across injury severity cohorts (Figure 3). Time from injury to baseline assessment and time between baseline assessment and last follow-up varied (Table 2) and were negatively and positively associated with the rate of change in mild and moderate-severe TBI respectively. We retained these variables as covariates in models.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInsert Figure 2\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eMachine learning modeling\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eResults of three supervised machine-learning algorithms, Random Forest (RF), Gradient Boosting (GB) and extreme GB (XGBoost), including both overall cognitive rate of change and domain-specific outcomes, respectively, can be seen in Figures 3, 5 and 7, for mild and moderate-severe TBI.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFeature importance heatmaps of each of the models are presented in Figure 3a and b, for mild and moderate-severe TBI, respectively. \u0026nbsp; For mild TBI, across the three models, the top PROGRESS-Plus feature was age, followed by education SD and the contextual parameter of GII. For moderate-severe TBI cohorts, the top PROGRESS-Plus feature was age SD and time related variables.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInsert\u0026nbsp;Figure 3\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eModel evaluation and interpretation\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eModel performance metrics are summarised in Table 2. Across all algorithms, predictive accuracy was comparable, with XGBoost slightly better in mean absolute error (MAE), and RF slightly better in root mean squared error (RMSE) in mild and moderate-severe TBI cohorts (Table 2) \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInsert Table 2\u003c/p\u003e\n\u003cp\u003eThe pattern of tuned hyperparameters (Table 3) across algorithms suggested different levels of structure in data in both mild and moderate-severe TBI. We observed that RF for both injury severity strata captured very shallow trees (i.e., maximum depth = 2) with the complexity in predictive signals explained by low-order interactions. The GB model favoured deeper individual trees (i.e., maximum depth of 9 and 6 for mild and moderate-severe TBI, respectively) where each successive tree focused on residual errors, allowing more complex interaction within a single tree compared to RF. The mild TBI model highlighted a very low learning rate (i.e., 0.01), with deeper trees and more iterations. The moderate-severe model used a higher learning rate (i.e., 0.1) with fewer trees and shallower depths. The XGBoost model, across mild and moderate-sever TBI strata, selected deeper trees using additional regularization to tolerate deeper depths without overfitting. The learning rate remained small, and overall complexity was determined by interactions with a number of boosting runs (Table 3). To support interpretation beyond hyperparameters, we examined variable influence and marginal effects (e.g., partial dependence/SHAP) to understand how each algorithm captured non-linear relationships and interactions in the data (Figure 4).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInsert Table 3\u003c/p\u003e\n\u003cp\u003eFor prediction, the Shapley Additive Explanations (SHAP) ranking showed that across models, time interval, GII, education (picked by RF), and age emerged as the strongest predictors of standardised rate of change for mild TBI. \u0026nbsp;For moderate-severe TBI cohorts, across models, baseline assessment, time interval, age, and GII (picked by RF), emerged as the strongest predictors of standardised rate of change (Figure 4). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInsert Figure 4\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eInternal validation and sensitivity analyses \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo ensure the robustness of the results, we performed 5-fold cross-validation to verify generalisability across folds (Table 4). For both injury severities, XGBoost showed slightly superior performance for standardised rate of change across data points. \u0026nbsp;The effect of all the primary predictors, covariates, and the effect modifiers remained the same (Supplement figures 4 and 5).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInsert Table 4\u003c/p\u003e\n\u003cp\u003eFurther analyses using a subset of data on executive function and learning and memory domains of cognition, which had relatively balanced data between mild and moderate-severe TBI cohorts, confirmed our hypothesis (i.e., that PROGRESS-Plus parameters would exert a more pronounced effect on domain-specific cognitive outcomes than on overall cognition, stronger for mild than moderate-severe TBI). Results showed that all the values of relative contributions for PROGRESS-Plus and time variables varied across models, however they remained largely the same as in the main analyses (Supplement Figures 6 and 10 for executive function and learning and memory TBI cohorts, respectively). The top three features for prediction of rate of change in executive function domain in mild TBI cohorts identified by the SHAP analyses were age, time interval, and education SD. These results were similar to the main analysis, with the exception of GII whose predictive value was reduced compared to the main analysis. In moderate-severe TBI cohorts, the top three predictive features remained as time interval, time from injury to baseline assessment, and age SD (Supplement Figure 7).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe top three consistent features for prediction of rate of change in the learning and memory domain in mild TBI cohorts identified by the SHAP analyses were education SD, time interval, and GII. These results were similar to the main analysis, with the exception of age whose predictive value was reduced compared to the main analysis. In moderate-severe TBI cohorts, the top three predictive features remained as time interval, time from injury to baseline assessment, and age SD, similar to the main analysis (Supplement Figure 11). Model performance for both domains remained similar to that of the main analysis (Supplement Table 4).\u003c/p\u003e\n\u003cp\u003eThe results of sensitivity analyses using 5-fold cross-validation as an internal validation approach to evaluate the robustness and generalizability of model estimates for specific domains of cognition are presented in Supplement Figures 8-9 and 12-13. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe targeted elimination of multicollinear age and education SD parameters provided further validity regarding the impact of time interval, age, and GII in mild TBI cohorts. In moderate-severe TBI, time related variables were further validated (Supplement Figures 14 and 15).\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we report innovative explainable ML research using published longitudinal data on the course of cognition after TBI, investigating PROGRESS-Plus characteristics of study samples as predictors of rate of change in cognition after injury. The results support the importance of considering social parameters in post-injury outcomes, and the ability of ML to explain heterogeneity\u003csup\u003e\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e in the course of cognition overall and by cognitive domain following TBI of equal severity. Our study has research, clinical, and policy implications.\u003c/p\u003e \u003cp\u003eWe found that among the most important predictors of cognitive change after TBI in published research across domains and injury severities were age, time-related parameters (i.e., time from injury to baseline assessment and time interval from baseline to last follow-up assessment), and country-level structural indicators. Given the pronounced variation in age and age SD across cohorts, the discussion brings attention to how researchers deal with age in their analyses in longitudinal studies. This is true for both mild and moderate-severe TBI.\u003c/p\u003e \u003cp\u003eAge-related deterioration affects all people, and is reflected across domains of cognition, reported for both mild and moderate-severe TBI cohorts\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e,\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e. The mechanisms that mediate age-related processes and changes reflected in the test performance of cohorts with TBI included in this study are not entirely clear; however, these processes are likely to be influenced by injury severity (Supplement Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003eb, e, f). Our results suggest that aging processes may be accelerated to a greater extent after moderate\u0026ndash;severe TBI as compared to mild, highlighting that neurorecovery and neurodegeneration occur simultaneously and therefore the interval between assessments does not carry the same meaning across injury severity cohorts. We found a higher rate of change in cognitive outcomes in moderate\u0026ndash;severe TBI (0.23) than in mild TBI (0.11), highlighting the value of how machine learning in capturing temporal effects in predicting outcomes. Our results also align with the random damage theories\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e,\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e, which situate around the disturbed balance between ongoing damage and repair that occurs in the process of natural aging, reflected in cognitive test performance. When the process is disturbed by brain injury, the balance of ongoing damage and repair is further dysregulated, and therefore was picked by ML as key predictors of the rate of change, to a greater extent in moderate-severe TBI than mild (Supplement Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003eb, e, f). The age of research participants, and the SD of the cohort, are, therefore expected to be seen, as it would affect the processes of both natural ageing and constrict recovery after injury, which is not expected to be uniform in participants of various ages. In addition, participants of different age in the cohort, reflected in the age SD of the cohort, are also more or less likely to suffer from multiple diseases associated with aging, which were rarely reported on in published research we included in the dataset, and which we were not able to test as predictors of the rate of change. Several longitudinal studies have uncovered that the age-specific strain caused by the onset of TBI in the presence of comorbidity is associated with cognitive outcomes, regardless of TBI severity\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. Future research should investigate age and age-related effects precisely to understand their predictive role in prognosis.\u003c/p\u003e \u003cp\u003eTime-related variables (i.e., time from injury to baseline assessment and time interval from baseline to last follow-up assessment, Supplement Table \u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e2\u003c/span\u003e), emerged as important predictors of rate of change, and therefore emphasize the critical role of time in prognostic research. We observed that the rate of cognitive change after TBI differed between mild and moderate-severe cohorts (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e). This finding aligns with prior research and with current evidence defining severe TBI as a risk factor for long-term cognitive decline\u003csup\u003e\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e,\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e. This study provides evidence that time affects rate of change across executive function and learning and memory domains of cognition, but we did not have sufficient data to test the effect on other cognitive domains (i.e., language, perceptual motor, complex attention, information processing speed, social cognition). Because our prior systematic review\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e and observational studies\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e,\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e brought attention to the fact that the course of cognition and recovery were not uniform and were dependent on the baseline assessment, future research should consider the implications of time after injury and baseline assessment in prognosis of cognitive domain-specific risks following injury.\u003c/p\u003e \u003cp\u003eOur results on relevance of baseline assessments (i.e., time zero) in prognosis were especially pronounced in moderate-severe TBI cohorts, with the mean baseline assessment conducted at around one year post injury, with great heterogeneity between cohorts (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e and dataset). In mild TBI, the mean baseline assessment was less than two weeks, with more homogeneity among samples (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e). At the point of baseline assessment, many participants in the mild TBI cohort may have reached a recovery plateau, which may not be the case for moderate-severe TBI, and therefore the time effects were not as pronounced in ML prediction for mild as in moderate-severe cohorts. Our results in sensitivity analyses confirmed the robustness of time effects, which underscore a critical need for coordinated research and policy efforts that explicitly integrate discussion on timing of research concerning prognosis based on injury severity, as this would impact scientific evidence with policy and practice implications. This is especially important as in both our prior hypothesis-driven approaches, we faced challenges in characterizing these temporal effects. Given that recovery occurs in parallel with processes of aging, these processes may reinforce one another with differing speed based on time that emerged from injury. Our results underscore that time is not merely a methodological detail, but a determinant of predictive accuracy and scientific interpretation. Future research should consider the benefits of ML for the prediction of outcomes after TBI with greater precision to time.\u003c/p\u003e \u003cp\u003eCountry-level indicators also warrant discussion. We systematically integrated heterogeneous datasets across multiple published cohorts on all published longitudinal research concerning the course of cognition after TBI. We found that ML captured GII as a predictor of the rate of change across several models in both injury severities in both bivariate and multivariate ML models. The positive correlation between GII and rate of change was consistent with previous reports of the important role of gender equality in heath outcomes\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e. These results are consistent with recent observations showing that trust in relationship are associated with outcomes in advanced clinical conditions, including stroke\u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e and dementia \u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e. In countries with social constraints on education and development, captured via GII and reflected in economic imbalances and restricted decision making, the macro-level relational environment may impose strains on both family and community relationships, impacting trust. While this may be perceived as far-fetched, results bring attention to the critical role of macro-level social parameters within existing data hierarchies and raise fundamental questions about their relative importance compared with person-level predictors. Greater emphasis on social pathways in future research will be essential to elucidate the mechanisms through which these parameters exert their effects on cognitive recovery.\u003c/p\u003e \u003cp\u003eOur correlation matrices results indicate that PROGRESS-Plus parameters characterising study samples, including whether the country of study origin is predominantly English speaking, gender/sex, education, and age are associated with rate of change. However, when these variables were used in ML, in consideration with other parameters (i.e., time effects and country-level structural indicators) their predictive values were not as strong as those of other parameters, except education and age. We have observed implications of the country-level structural indicators expressed by GII in results of our past systematic review\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e, where when we positioned results along the GII continuum, differences between male and females across outcomes and countries of publication started to dissipate. The GII is a composite, time-varying indicator of a country\u0026rsquo;s gender inequality level, incorporating measures of educational attainment, labour force participation, maternal mortality, adolescent fertility, and parliamentary representation\u003csup\u003e\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e. Therefore, when the index was used in ML models, the predictive effect of cohorts\u0026rsquo; education and education SD were diminished, and several of our models featured the GII as a salient predictor of the rate of cognitive change after TBI, in both mild and moderate-severe samples. While it has been previously suggested that cognitive outcomes after TBI differ between people of different gender/sex and education level, our results are the first to highlight that the importance of these parameters are affected by broader social and structural contexts. This has important research, clinical, and policy implications, particularly in ongoing debates regarding the long-term cognitive consequences of milder forms of TBI, and the cumulative contributions of biological sex, gender, education, and injury severity to TBI outcomes. Our ML results also bring attention to the intersectionality reflected in the GII, which may also have implications for those who participated in research included in the dataset, and therefore was captured by ML as a prognostic parameter after TBI. Evidence increasingly shows that the characteristics of research participants in TBI studies are not truly reflective of TBI populations, where racialized, less educated, and more disadvantaged communities are less like to participate in research but are also disproportionately affected by TBI and its adverse consequences \u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e,\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e,\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e. Future prognostic research and policy should diversify research samples and strive for systematic inclusion of diverse groups of people in TBI research. Overcoming barriers to research remains a priority\u003csup\u003e\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e,\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe purpose of our research was to delineate whether social parameters associated with rate of change of cognition using data from published TBI cohorts, and to provide information on whether the characteristics of longitudinal study samples play a role in prognosis. We applied three ML approaches to build prognostic models concerning rate of cognitive change in both mild and moderate-severe injury severity samples. All three ML approaches converge on the conclusion that social and structural characteristics of published cohorts are prognostically meaningful for cognitive change after TBI. Models for mild TBI strata were more heterogeneous and prone to social influences compared to moderate-severe TBI strata, highlighted in the complexity reflected in the tines hyperparameters of boosting models (GB and XGBoost). Nonetheless, all models indicated that sample compositions are determining prognostic patterns and therefore should be treated with greater care in future research. Results of ML models not only delineated four PROGRESS-Plus characteristics (i.e., country of English language dominance as place of residence, gender/sex, education, and age) of cohorts as important determinants of prognosis, but also exposed critical gaps in the current prognostic evidence base by showing who contributes to longitudinal research and which social variables are measured in cohort studies on prognosis and which are not (i.e., religion/spirituality, socioeconomic status, occupation, social capital), constraining what can be known about the value of these parameters to the course of cognition and prognosis after TBI. Therefore, ML approaches as a methodological tool in research is of great value not only for improving prognostic modelling and meta-research, but also to bring attention to how methodological choices made by researchers about sampling\u003csup\u003e\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e, anchored time scale (i.e., time zero and elapsed time)\u003csup\u003e\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e\u003c/sup\u003e, and cognitive measures\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e that actively shaped the current best state of knowledge available to clinicians and policymakers.\u003c/p\u003e \u003cp\u003eOne limitation of our study is limited reporting of study participants\u0026rsquo; social profiles in the dataset of published studies. We developed and implemented rigorous data harmonization processes to maximise the ability to examine all available PROGRESS-Plus characteristics in the dataset prior to application of ML-based prognostic modelling. Despite our data harmonization efforts, the dataset exhibited substantial missingness across several PROGRESS-Plus parameters, including race/ethnicity, socioeconomic status, occupation, religion/spirituality, and social capital. Although data augmentation using synthetic data generation is sometimes employed to mitigate this issue in ML, implementation of this approach was not possible in the present study, as these PROGRESS-Plus parameters were sparsely reported, and the available data was less than 22% for occupation and less than 5% for all other parameters after data harmonization (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e). At such low levels of representation, synthetic data generation would have relied on insufficient underlying distributions, increasing the risk of amplifying noise, introducing artificial patterns, and reinforcing existing biases. For these reasons, we restricted model training to four parameters with adequate representation: the country of study origin through country of English language dominance, gender/sex, education, and age. Future research should consider this limitation and commit to a standardised reporting of PROGRESS-Plus characteristics to enable equitable and robust ML-based prognostic modelling using published data. Further investigations are also needed to refine the role of PROGRESS-Plus parameters as predictors of outcomes when considering structural level parameters.\u003c/p\u003e \u003cp\u003eOur research and results provide a foundation for advancing prediction models that systematically evaluate how study sample representativeness affect TBI prognosis. We employed three ML models, including the RF, the GB, and the XGBoost. Employing SHAP in the present study provided additional insights into model performance and predictive value of each parameter in the presence of others, in both mild and moderate-severe injury severity study samples. The usability of the models in predicting outcomes using published data depends on data quality. We recommend that future researchers and who will apply our process and work with published data dedicate significant time to data extraction, pre-processing and standardization prior to machine learning model input. In addition, users are expected to have at least some machine learning expertise and be familiar with the variables in the dataset, as well as with the principles of modelling. While this may limit direct use by non-technical personnel, it ensures reliable and accurate predictions for research and policy applications. Future enhancements in prognosis research that can be used in machine learning require the systematic collection, reporting, and integration of social parameters in research to enable more equitable clinical and policy decision-making.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eWe conducted and reported our research following the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis \u0026ndash; Artificial Intelligence (TRIPOD-AI) guidelines for prediction model development and validation\u003csup\u003e\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e\u003c/sup\u003e. We shared all research-related files in an open data repository on the Open Science Framework (OSF)\u003csup\u003e\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e\u003c/sup\u003e, in alliance with the FAIR Guiding Principles\u003csup\u003e\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e\u003c/sup\u003e. The complete dataset is available upon reasonable request from the corresponding author.\u003c/p\u003e\n\u003ch3\u003eData sources and description\u003c/h3\u003e\n\u003cp\u003eThe dataset used in this work was developed by our research team using data from our prior systematic reviews\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e on the course and predictors of cognitive outcomes after TBI. We updated this dataset with new evidence that emerged since the original searches\u003csup\u003e\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e\u003c/sup\u003e. The flow diagram of the study selection is depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003e. In brief, the data covers evidence that emerged from five databases: Embase, Medline, Scopus, and Cochrane Central Register for Controlled Trials, and PsycINFO. We searched the databases from inception until April 8, 2024 for peer-reviewed English-language longitudinal studies reporting raw cognitive test scores at two or more time points in adults diagnosed with TBI. Only studies that provided data on the course of cognition over at least two data points after TBI were included. The protocol, PROSPERO registry, and the systematic reviews that provided specifics on study methodology and the dataset are published\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e,\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e,\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eOur research concerned data on TBI diagnosis, injury severity and cause of injury data, and PROGRESS-Plus parameters of each study sample (country of recruitment, race/language/ethnicity, occupation, gender/sex, religion/spirituality, education, socioeconomic status, social capital, age, and other contextual parameters), and cognitive test scores at baseline and last follow-up assessments\u003csup\u003e\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e\u003c/sup\u003e. Each cognitive test in a given cohort was treated as a unit of analysis.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eDataset preparation and machine learning workflow\u003c/h2\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003e represents dataset preparation and machine learning workflow, outlining seven steps: (i) data collection; (ii) data extraction; (iii) data harmonization; (iv) data preprocessing; (v) machine learning algorithms; (vi) model evaluation; and (vii) sensitivity analysis, which allowed generation of overall and model-specific explanations.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eData collection\u003c/h2\u003e \u003cp\u003eStudy selection followed predefined inclusion and exclusion criteria. Eligible studies included those that reported longitudinal data on standardised cognitive tests and provided sufficient injury-related data of study sample characteristics to support data harmonization and modeling. For more information, please refer to published studies\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eData extraction\u003c/h2\u003e \u003cp\u003eTwo independent researchers extracted data using standardized, previously published extraction forms\u003csup\u003e\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e\u003c/sup\u003e. Extracted variables included demographic characteristics (age, sex/gender, race/ethnicity), social parameters (place of residence, occupation, education, and social capital), and contextual brain injury-related parameters (injury severity, mechanism of injury, comorbidities, and time from injury to baseline assessment). We also extracted data on time from injury to baseline assessment, and calculated the time interval between baseline and last follow-up cognitive assessments. Data was extracted from the main text and supplementary materials of each published study by two independent researchers. Extraction accuracy and reliability were ensured through double extraction, cross-checking, and resolution of discrepancies by consensus.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eData harmonization process\u003c/h2\u003e \u003cp\u003eWe considered using ML-based solutions for data harmonization of several variables\u003csup\u003e\u003cspan additionalcitationids=\"CR70\" citationid=\"CR69\" class=\"CitationRef\"\u003e69\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e71\u003c/span\u003e\u003c/sup\u003e. However, we faced significant challenges in finding scientific guidance on the application of standardized ML data harmonization in practice. We found that each harmonization case was a unique process and necessitated a series of context-dependent decisions through iterative team discussions where we evaluated several harmonization options at once (i.e., flexible harmonization\u003csup\u003e\u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e72\u003c/span\u003e,\u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e73\u003c/span\u003e\u003c/sup\u003e).\u003c/p\u003e \u003cp\u003eWe first created a library of terms based on all extracted data (Supplement Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e2\u003c/span\u003e) that defined variable names, operational definitions, categories, and units to ensure consistency across included cohorts. The core team discussed each variable in the table and evaluated different variables under the same section header, jointly deciding how to incorporate these updates into the harmonized protocol template. We then harmonised data to reconcile different operationalisations of similar concepts (i.e., gender/sex composition of study samples, reported in a variety of data formats, known as syntax) as well as semantics (i.e., intended meanings of words such as young adults, primary schooling, etc.), conceptual schema (i.e., structured and unstructured data extracted from tables and raw text, respectively), and measurement differences (e.g. when the same concept was reported with different measurements, including injury severity, cognitive domain, etc.). Finally, we transformed data to numeric values in order to prepare the data for ML (Supplement Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e3\u003c/span\u003e). We structured extracted social and contextual elements using the PROGRESS-Plus framework. The categorical parameters of study samples were transformed to continuous variables where feasible: gender/sex composition was transformed to proportion men/males, and education level was transformed to education in years. Additionally, we transformed age-related parameters to continuous age and age standard deviation of study samples. Country of study origin (where recruitment took place) was transformed to country of English language dominance as a binary variable (0, 1).\u003c/p\u003e \u003cp\u003eCognitive outcome scores of standardised tests at baseline and last follow-up and the domains of cognition they reflect were represented as mean and standard deviation, and binary values, respectively.\u003c/p\u003e \u003cp\u003eTo ensure quality, data harmonization and transformation processes were conducted by two independent researchers. We documented the process and decisions (Supplement Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e3\u003c/span\u003e) made for future data users in compliance with the FAIR standards\u003csup\u003e\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eData preprocessing\u003c/h2\u003e \u003cp\u003eWe conducted data preprocessing. This step included scaling of continuous variables, systematic assessment of outliers, and correction for imbalances across injury severity groups based on visual plots and correlation matrices between predictors and the outcome.\u003c/p\u003e \u003cp\u003eWhen time was reported categorically or as ranges, we used midpoint values. For cohorts reporting multiple follow-up time points\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e, the final available follow-up assessment was used for data harmonization.\u003c/p\u003e \u003cp\u003eWe calculated time SD for each data point, considering cohort sample size and reported outcome score, and score SDs in order to preserve the variance structure across included cohorts. We then calculated rate of change per month as the outcome for longitudinal harmonisation of data points. For each data point, the rate of change was computed by dividing the difference between outcome values at baseline and last follow-up by the elapsed time between assessments using the following formula:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\text{R}\\text{a}\\text{t}\\text{e}\\:\\text{o}\\text{f}\\:\\text{c}\\text{o}\\text{g}\\text{n}\\text{i}\\text{t}\\text{i}\\text{v}\\text{e}\\:\\text{c}\\text{h}\\text{a}\\text{n}\\text{g}\\text{e}\\:\\left(\\text{p}\\text{e}\\text{r}\\:\\text{m}\\text{o}\\text{n}\\text{t}\\text{h}\\right)=\\frac{{Y}_{FU-}{Y}_{BL}}{T/30}\\mathbf{}\\mathbf{}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere Y\u003csub\u003eFU\u003c/sub\u003e denotes the cognitive test score at the last follow-up assessment, Y\u003csub\u003eBL\u003c/sub\u003e denotes the cognitive test score at baseline, and T denotes the mean follow-up interval measured in days. The follow-up interval was converted to months by dividing by 30 to harmonize time scales across studies.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eClass imbalance\u003c/h2\u003e \u003cp\u003eWe observed that imbalance in our data occurred within injury severity categories. To mitigate the influence of this imbalance on model performance and to preserve clinically meaningful distinctions, we stratified analyses by injury severity rather than applying resampling-based class-imbalance methods. Moderate, moderate\u0026ndash;severe, and severe injuries were combined into a single \u0026lsquo;moderate\u0026ndash;severe\u0026rsquo; category, and models were trained separately for mild and moderate\u0026ndash;severe TBI.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eMissing data\u003c/h2\u003e \u003cp\u003eWe made a priori decision to restrict ML analyses to complete cases, where only data points with complete information on the variables required for ML modelling were included in the final analytic sample\u003csup\u003e\u003cspan additionalcitationids=\"CR75\" citationid=\"CR74\" class=\"CitationRef\"\u003e74\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e76\u003c/span\u003e\u003c/sup\u003e. This decision was made because many variables expressed structurally missing data reporting (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e2\u003c/span\u003e) rather than missing at random.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eMachine learning modeling\u003c/h2\u003e \u003cp\u003eWe applied three supervised ML algorithms\u003csup\u003e\u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e77\u003c/span\u003e\u003c/sup\u003e to mild and moderate-severe injury severity datasets: random forest (RF)\u003csup\u003e\u003cspan citationid=\"CR78\" class=\"CitationRef\"\u003e78\u003c/span\u003e,\u003cspan citationid=\"CR79\" class=\"CitationRef\"\u003e79\u003c/span\u003e\u003c/sup\u003e, gradient boosting (GB)\u003csup\u003e\u003cspan citationid=\"CR80\" class=\"CitationRef\"\u003e80\u003c/span\u003e,\u003cspan citationid=\"CR81\" class=\"CitationRef\"\u003e81\u003c/span\u003e\u003c/sup\u003e, and extreme gradient boosting (XGBoost)\u003csup\u003e\u003cspan citationid=\"CR82\" class=\"CitationRef\"\u003e82\u003c/span\u003e\u003c/sup\u003e, to model cognitive change over time. These models were trained to predict both overall cognitive trajectories (standardised rate of change per month) and domain-specific standardised rate of change using harmonised PROGRESS-Plus variables. To ensure comparability across algorithms and mitigate bias introduced by differences in cohorts\u0026rsquo; age, time between injury and baseline assessment, and follow-up times, all models were trained on harmonised rates of change per unit of time, and controlled for time between TBI and baseline assessment and time between assessments.\u003c/p\u003e \u003cp\u003eWe considered time from injury to baseline assessment, time between baseline assessment and last follow-up, and country-level structural indicators as potential modifiers of the relationship between primary predictors and cognitive outcomes. This decision was made in order to capture historical evolution of populations\u0026rsquo; social parameters by country, including the GII and HDI.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eModel evaluation and interpretation\u003c/h2\u003e \u003cp\u003eWe evaluated model performance using mean absolute error (MAE) and root mean square error (RMSE)\u003csup\u003e\u003cspan citationid=\"CR83\" class=\"CitationRef\"\u003e83\u003c/span\u003e\u003c/sup\u003e. We then examined feature importance using feature importance heatmaps and the Shapley Additive Explanations (SHAP)\u003csup\u003e\u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e84\u003c/span\u003e\u003c/sup\u003e. By observing SHAP values, we evaluated the predictive importance of each feature and how it contributes to the difference between an actual prediction and a mean prediction, analyzing non-linear relationships between PROGRESS-Plus variables and rate of change.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eInternal validation and sensitivity analysis\u003c/h2\u003e \u003cp\u003eWe performed sensitivity analyses using 5-fold cross-validation as an internal validation approach to evaluate the robustness and generalizability of model estimates. This procedure assessed the stability of the effects of input parameters on cognitive outcomes across injury severity strata.\u003c/p\u003e \u003cp\u003eWe also conducted sensitivity analyses to assess the potential effects of PROGRESS-Plus parameters on specific cognitive domains that provided sufficient data and had relatively even distribution between mild and moderate-severe TBI cohorts. We hypothesized a priori that PROGRESS-Plus parameters would exert a more pronounced effect on domain-specific cognitive outcomes than on overall cognition, stronger for mild than moderate-severe TBI, given previously observed variability in rates of change by injury severity within the cohorts\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. This allowed us to verify the relative contributions of the important PROGRESS-Plus parameters under a more constrained, cognitive domain-specific dataset. We re-evaluated feature importance using feature importance heatmaps and the SHAP, estimating each predictor\u0026rsquo;s contribution to standardised rate of cognitive change in cognitive domain-specific datasets for executive function and learning and memory domains of cognition.\u003c/p\u003e \u003cp\u003eWe also investigated the model performance by removing features with high correlation coefficients to evaluate the impact of multicollinearity on the model\u0026rsquo;s performance. By selectively excluding highly correlated features (education and age SD), we aimed to improve the robustness of the model.\u003c/p\u003e \u003cp\u003eAll analyses, including generation of figures, were performed using Python (version 3)\u003csup\u003e85,86\u003c/sup\u003e. Figures were created using Python libraries (matplotlib (v3.x), seaborn (v0.x), and wordcloud (v1.x)).\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003eThe study protocol was approved by the ethics committees at the clinical (Toronto Rehabilitation Institute\u0026ndash;University Health Network) and academic (University of Toronto) institutions. All methods were carried out in accordance with the relevant guidelines and regulations. Informed consent: This research utilised published data with no access to personal information.\u003c/p\u003e\u003cp\u003e \u003cstrong\u003eCompeting Interests Statement\u003c/strong\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eEthical approval \u003cb\u003eand informed consent\u003c/b\u003e \u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eInformed consent:\u003c/h2\u003e \u003cp\u003eThis research utilised published data with no access to personal information.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eCIHR Brain Health and Reduction of Risk for Age-related Cognitive Impairment SGD #202306BK5-510306-BKS-ADHD-220229, and Canada Research Chair in Neurological Disorders and Brain Health, CRC-2021-00074.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eTM and TTS conceived the project. TM supervised US and TTS data extraction and harmonization. TM developed templates for data library encoding and transformation processing. JX and US performed the data analyses with methodological supervision from TM. US created visual display of data. TM wrote the first draft of the manuscript with input from US. All authors contributed to data interpretation and editing of the paper.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003e The authors thank Professor Michael Escobar for contribution to the technical discussion on mathematical evolution of the concept; trainees of the BRIDGE Lab (bridgelab.ca), Ashlee Kim, Chuxi Pan, and Mursal Jahed, for their support with title and abstract screening and data extraction; and Information Specialists from Library Services at the University Health Network (Jessica Babineau, Emilia Main, and Cynthia Chui) for help in retrieving data for systematic reviews that formed the basis for the data used in this research.This study was funded by the Canadian Institutes of Health Research (CIHR) Brain Health and Reduction of Risk for Age-related Cognitive Impairment SGD #202306BK5-510306-BKS-ADHD-220229, and in part by Canada Research Chair in Neurological Disorders and Brain Health, CRC-2021-00074. This research program was developed with endorsement by patient and public organizations, who reviewed the research program prior to its submission to the funding agency (Canadian Institute of Health Research). The critical need to address PROGRESS-Plus parameters in brain injury research was also endorsed by team members of the SGD #202306BK5-510306-BKS-ADHD-220229. The content is solely the authors' responsibility and does not necessarily represent the official views of the CIHR.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data encoding and harmonisation templates used in this research are available from the public [data portal](https:/osf.io/jku23/overview) . The full dataset creation plan and underlying analytic code are available from the corresponding author upon request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMollayeva, T., Mollayeva, S., Pacheco, N., D\u0026rsquo;Souza, A. \u0026amp; Colantonio, A. The course and prognostic factors of cognitive outcomes after traumatic brain injury: A systematic review and meta-analysis. \u003cem\u003eNeurosci. Biobehav Rev.\u003c/em\u003e \u003cb\u003e99\u003c/b\u003e, 198\u0026ndash;250 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTritt, A. et al. Data-driven distillation and precision prognosis in traumatic brain injury with interpretable machine learning. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 21200 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBark, D. et al. Refining outcome prediction after traumatic brain injury with machine learning algorithms. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 8036 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOldenburg, C., Lundin, A., Edman, G., Nygren-de Boussard, C. \u0026amp; Bartfai, A. Cognitive reserve and persistent post-concussion symptoms\u0026mdash;A prospective mild traumatic brain injury (mTBI) cohort study. \u003cem\u003eBrain Inj\u003c/em\u003e. \u003cb\u003e30\u003c/b\u003e, 146\u0026ndash;155 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTheadom, A. et al. Social cognition four years after mild-TBI: An age-matched prospective longitudinal cohort study. \u003cem\u003eNeuropsychology\u003c/em\u003e \u003cb\u003e33\u003c/b\u003e, 560\u0026ndash;567 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHANAFY, S. et al. Comorbidity in traumatic brain injury and functional outcomes: a systematic review. \u003cem\u003eEur J. Phys. Rehabil Med\u003c/em\u003e \u003cb\u003e57\u003c/b\u003e, (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eB\u0026uuml;ttner, F., Terry, D. P. \u0026amp; Iverson, G. L. Using a Likelihood Heuristic to Summarize Conflicting Literature on Predictors of Clinical Outcome Following Sport-Related Concussion. \u003cem\u003eClin. J. Sport Med.\u003c/em\u003e \u003cb\u003e31\u003c/b\u003e, e476\u0026ndash;e483 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMollayeva, T., Mollayeva, S., Pacheco, N. \u0026amp; Colantonio, A. Systematic Review of Sex and Gender Effects in Traumatic Brain Injury: Equity in Clinical and Functional Outcomes. \u003cem\u003eFront Neurol\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcIntosh, S. J., Vergeer, M. H., Galarneau, J. M., Eliason, P. H. \u0026amp; Debert, C. T. Factors Associated With Persisting Symptoms After Concussion in Adults With Mild TBI. \u003cem\u003eJAMA Netw. Open.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, e2516619 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMollayeva, T. et al. Decoding health status transitions of over 200 000 patients with traumatic brain injury from preceding injury to the injury event. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 5584 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMollayeva, T., Tran, A., Chan, V., Colantonio, A. \u0026amp; Escobar, M. D. Sex-specific analysis of traumatic brain injury events: applying computational and data visualization techniques to inform prevention and management. \u003cem\u003eBMC Med. Res. Methodol.\u003c/em\u003e \u003cb\u003e22\u003c/b\u003e, 30 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeterina, A. et al. Gender versus sex in predicting outcomes of traumatic brain injury: a cohort study utilizing large administrative databases. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 18453 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYaniv, D. Trust the Process: A New Scientific Outlook on Psychodramatic Spontaneity Training. \u003cem\u003eFront Psychol\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHanazaki, N. Local and traditional knowledge systems, resistance, and socioenvironmental justice. \u003cem\u003eJ. Ethnobiol. Ethnomed.\u003c/em\u003e \u003cb\u003e20\u003c/b\u003e, 5 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim, H., Jahed, M., Sant\u0026rsquo;Ana, T. T. \u0026amp; Mollayeva, T. (eds) (Ashlee), Cognitive Outcomes Among Adults Following Traumatic Brain Injury: A Systematic Review Using the PROGRESS-plus Framework. \u003cem\u003eArch Phys Med Rehabil\u003c/em\u003e 106, e120 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBleiberg, J. et al. Duration of Cognitive Impairment After Sports Concussion. \u003cem\u003eNeurosurgery\u003c/em\u003e \u003cb\u003e54\u003c/b\u003e, 1073\u0026ndash;1080 (2004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, C. C., Wei, S. T., Tsaia, S. C., Chen, X. X. \u0026amp; Cho, D. Y. Cerebrolysin enhances cognitive recovery of mild traumatic brain injury patients: double-blind, placebo-controlled, randomized study. \u003cem\u003eBr. J. Neurosurg.\u003c/em\u003e \u003cb\u003e27\u003c/b\u003e, 803\u0026ndash;807 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCovassin, T., Stearne, D. \u0026amp; Elbin, R. Concussion History and Postconcussion Neurocognitive Performance and Symptoms in Collegiate Athletes. \u003cem\u003eJ. Athl Train.\u003c/em\u003e \u003cb\u003e43\u003c/b\u003e, 119\u0026ndash;124 (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDikmen, S., Machamer, J. \u0026amp; Temkin, N. Mild Traumatic Brain Injury: Longitudinal Study of Cognition, Functional Status, and Post-Traumatic Symptoms. \u003cem\u003eJ. Neurotrauma\u003c/em\u003e. \u003cb\u003e34\u003c/b\u003e, 1524\u0026ndash;1530 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarbota, K. D. et al. Longitudinal diffusion tensor imaging and neuropsychological correlates in traumatic brain injury patients. \u003cem\u003eFront Hum. Neurosci\u003c/em\u003e \u003cb\u003e6\u003c/b\u003e, (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eField, M., Collins, M. W., Lovell, M. R. \u0026amp; Maroon, J. Does age play a role in recovery from sports-related concussion? A comparison of high school and collegiate athletes. \u003cem\u003eJ. Pediatr.\u003c/em\u003e \u003cb\u003e142\u003c/b\u003e, 546\u0026ndash;553 (2003).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHicks, A. J. et al. Does cognitive decline occur decades after moderate to severe traumatic brain injury? A prospective controlled study. \u003cem\u003eNeuropsychol. Rehabil\u003c/em\u003e. \u003cb\u003e32\u003c/b\u003e, 1530\u0026ndash;1549 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKersel, D. A., Marsh, N. V., Havill, J. H. \u0026amp; Sleigh, J. W. Neuropsychological functioning during the year following severe traumatic brain injury. \u003cem\u003eBrain Inj\u003c/em\u003e. \u003cb\u003e15\u003c/b\u003e, 283\u0026ndash;296 (2001).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKontos, A. P. et al. The effects of combat-related mild traumatic brain injury (mTBI). \u003cem\u003eJ. Trauma. Acute Care Surg.\u003c/em\u003e \u003cb\u003e79\u003c/b\u003e, S146\u0026ndash;S151 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKwok, F. Y., Lee, T. M. C., Leung, C. H. S. \u0026amp; Poon, W. S. Changes of cognitive functioning following mild traumatic brain injury over a 3-month period. \u003cem\u003eBrain Inj\u003c/em\u003e. \u003cb\u003e22\u003c/b\u003e, 740\u0026ndash;751 (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, H. et al. Comparing effects of methylphenidate, sertraline and placebo on neuropsychiatric sequelae in patients with traumatic brain injury. \u003cem\u003eHum. Psychopharmacology: Clin. Experimental\u003c/em\u003e. \u003cb\u003e20\u003c/b\u003e, 97\u0026ndash;104 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu, J., Roe, C., Sigurdardottir, S., Andelic, N. \u0026amp; Forslund, M. Trajectory of Functional Independent Measurements during First Five Years after Moderate and Severe Traumatic Brain Injury. \u003cem\u003eJ. Neurotrauma\u003c/em\u003e. \u003cb\u003e35\u003c/b\u003e, 1596\u0026ndash;1603 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacciocchi, S. N., Seel, R. T. \u0026amp; Thompson, N. The Impact of Mild Traumatic Brain Injury on Cognitive Functioning Following Co-occurring Spinal Cord Injury. \u003cem\u003eArch. Clin. Neuropsychol.\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e, 684\u0026ndash;691 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMalec, J. F. et al. Longitudinal Effects of Medical Comorbidities on Functional Outcome and Life Satisfaction After Traumatic Brain Injury: An Individual Growth Curve Analysis of NIDILRR Traumatic Brain Injury Model System Data. \u003cem\u003eJ. Head Trauma. Rehabilitation\u003c/em\u003e. \u003cb\u003e34\u003c/b\u003e, E24\u0026ndash;E35 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMandleberg, I. A. Cognitive recovery after severe head injury. 3. WAIS verbal and performance IQs as a function of post-traumatic amnesia duration and time from injury. \u003cem\u003eJ. Neurol. Neurosurg. Psychiatry\u003c/em\u003e. \u003cb\u003e39\u003c/b\u003e, 1001\u0026ndash;1007 (1976).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePar\u0026eacute;, N., Rabin, L. A., Fogel, J. \u0026amp; P\u0026eacute;pin, M. Mild traumatic brain injury and its sequelae: Characterisation of divided attention deficits. \u003cem\u003eNeuropsychol. Rehabil\u003c/em\u003e. \u003cb\u003e19\u003c/b\u003e, 110\u0026ndash;137 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobertson, K. \u0026amp; Schmitter-Edgecombe, M. Self-awareness and traumatic brain injury outcome. \u003cem\u003eBrain Inj\u003c/em\u003e. \u003cb\u003e29\u003c/b\u003e, 848\u0026ndash;858 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSandhaug, M., Andelic, N., Langhammer, B. \u0026amp; Mygland, A. Functional level during the first 2 years after moderate and severe traumatic brain injury. \u003cem\u003eBrain Inj\u003c/em\u003e. \u003cb\u003e29\u003c/b\u003e, 1431\u0026ndash;1438 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmitter-Edgecombe, M. \u0026amp; Robertson, K. Recovery of visual search following moderate to severe traumatic brain injury. \u003cem\u003eJ. Clin. Exp. Neuropsychol.\u003c/em\u003e \u003cb\u003e37\u003c/b\u003e, 162\u0026ndash;177 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSNOW, P., PONSFORD, J. \u0026amp; DOUGLAS, J. \u0026amp; Conversational discourse abilities following severe traumatic brain injury: a follow up study. \u003cem\u003eBrain Inj\u003c/em\u003e. \u003cb\u003e12\u003c/b\u003e, 911\u0026ndash;935 (1998).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVanderploeg, R. D., Donnell, A. J., Belanger, H. G. \u0026amp; Curtiss, G. Consolidation deficits in traumatic brain injury: The core and residual verbal memory defect. \u003cem\u003eJ. Clin. Exp. Neuropsychol.\u003c/em\u003e \u003cb\u003e36\u003c/b\u003e, 58\u0026ndash;73 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, S. et al. Umbilical cord mesenchymal stem cell transplantation significantly improves neurological function in patients with sequelae of traumatic brain injury. \u003cem\u003eBrain Res.\u003c/em\u003e \u003cb\u003e1532\u003c/b\u003e, 76\u0026ndash;84 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZaninotto, A. L. et al. Visuospatial memory improvement in patients with diffuse axonal injury (DAI): a 1-year follow-up study. \u003cem\u003eActa Neuropsychiatr\u003c/em\u003e. \u003cb\u003e29\u003c/b\u003e, 35\u0026ndash;42 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrigatano, G. P. et al. Neuropsychological rehabilitation after closed head injury in young adults. \u003cem\u003eJ. Neurol. Neurosurg. Psychiatry\u003c/em\u003e. \u003cb\u003e47\u003c/b\u003e, 505\u0026ndash;513 (1984).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSours, C. et al. Associations between interhemispheric functional connectivity and the Automated Neuropsychological Assessment Metrics (ANAM) in civilian mild TBI. \u003cem\u003eBrain Imaging Behav.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 190\u0026ndash;203 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStenberg, J. et al. Change in self-reported cognitive symptoms after mild traumatic brain injury is associated with changes in emotional and somatic symptoms and not changes in cognitive performance. \u003cem\u003eNeuropsychology\u003c/em\u003e \u003cb\u003e34\u003c/b\u003e, 560\u0026ndash;568 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWylie, G. R. et al. Cognitive Improvement after Mild Traumatic Brain Injury Measured with Functional Neuroimaging during the Acute Period. \u003cem\u003ePLoS One\u003c/em\u003e. \u003cb\u003e10\u003c/b\u003e, e0126110 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePowell, T. J., Collin, C. \u0026amp; Sutton, K. A follow-up study of patients hospitalized after minor head injury. \u003cem\u003eDisabil. Rehabil\u003c/em\u003e. \u003cb\u003e18\u003c/b\u003e, 231\u0026ndash;237 (1996).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBarker-Collo, S. et al. Neuropsychological outcome and its correlates in the first year after adult mild traumatic brain injury: A population-based New Zealand study. \u003cem\u003eBrain Inj\u003c/em\u003e. \u003cb\u003e29\u003c/b\u003e, 1604\u0026ndash;1616 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcCREA, M. et al. Standard regression-based methods for measuring recovery after sport-related concussion. \u003cem\u003eJ. Int. Neuropsychol. Soc.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 58\u0026ndash;69 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRajkomar, A., Hardt, M., Howell, M. D., Corrado, G. \u0026amp; Chin, M. H. Ensuring Fairness in Machine Learning to Advance Health Equity. \u003cem\u003eAnn. Intern. Med.\u003c/em\u003e \u003cb\u003e169\u003c/b\u003e, 866\u0026ndash;872 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMurman, D. The Impact of Age on Cognition. \u003cem\u003eSemin Hear.\u003c/em\u003e \u003cb\u003e36\u003c/b\u003e, 111\u0026ndash;121 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAyton, A., Spitz, G., Hicks, A. J. \u0026amp; Ponsford, J. Ageing with Traumatic Brain Injury: Long-Term Cognition and Wellbeing. \u003cem\u003eJ. Neurotrauma\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1089/neu.2024.0524\u003c/span\u003e\u003cspan address=\"10.1089/neu.2024.0524\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKirkwood, T. B. L. \u0026amp; Austad, S. N. Why do we age? \u003cem\u003eNature\u003c/em\u003e \u003cb\u003e408\u003c/b\u003e, 233\u0026ndash;238 (2000).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKirkwood, T. B. L. Evolution of ageing. \u003cem\u003eNature\u003c/em\u003e \u003cb\u003e270\u003c/b\u003e, 301\u0026ndash;304 (1977).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChan, V., Mollayeva, T., Ottenbacher, K. J. \u0026amp; Colantonio, A. Sex-Specific Predictors of Inpatient Rehabilitation Outcomes After Traumatic Brain Injury. \u003cem\u003eArch. Phys. Med. Rehabil\u003c/em\u003e. \u003cb\u003e97\u003c/b\u003e, 772\u0026ndash;780 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMollayeva, T., Hurst, M., Escobar, M. \u0026amp; Colantonio, A. Sex-specific incident dementia in patients with central nervous system trauma. \u003cem\u003eAlzheimer\u0026rsquo;s Dementia: Diagnosis Assess. Disease Monit.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 355\u0026ndash;367 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eD\u0026rsquo;Souza, A. et al. Measuring Change Over Time: A Systematic Review of Evaluative Measures of Cognitive Functioning in Traumatic Brain Injury. \u003cem\u003eFront Neurol\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBryant, A. M. et al. Profiles of Cognitive Functioning at 6 Months After Traumatic Brain Injury Among Patients in Level I Trauma Centers. \u003cem\u003eJAMA Netw. Open.\u003c/em\u003e \u003cb\u003e6\u003c/b\u003e, e2349118 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArmstrong, K. \u0026amp; Ritchie, C. Research Participation in Marginalized Communities \u0026mdash; Overcoming Barriers. \u003cem\u003eN. Engl. J. Med.\u003c/em\u003e \u003cb\u003e386\u003c/b\u003e, 203\u0026ndash;205 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarshall, I. J. et al. The effects of socioeconomic status on stroke risk and outcomes. \u003cem\u003eLancet Neurol.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 1206\u0026ndash;1218 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOrnish, D. et al. Effects of intensive lifestyle changes on the progression of mild cognitive impairment or early dementia due to Alzheimer\u0026rsquo;s disease: a randomized, controlled clinical trial. \u003cem\u003eAlzheimers Res. Ther.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e, 122 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGender Inequality Index (GII). \u003cem\u003eUnited Nations Development Program: Human Development Reports\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://hdr.undp.org/data-center/thematic-composite-indices/gender-inequality-index#/indicies/GII\u003c/span\u003e\u003cspan address=\"https://hdr.undp.org/data-center/thematic-composite-indices/gender-inequality-index#/indicies/GII\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCookson, R., Robson, M., Skarda, I. \u0026amp; Doran, T. Equity-informative methods of health services research. \u003cem\u003eJ. Health Organ. Manag\u003c/em\u003e. \u003cb\u003e35\u003c/b\u003e, 665\u0026ndash;681 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCastillo, E. G. \u0026amp; Harris, C. Directing Research Toward Health Equity: a Health Equity Research Impact Assessment. \u003cem\u003eJ. Gen. Intern. Med.\u003c/em\u003e \u003cb\u003e36\u003c/b\u003e, 2803\u0026ndash;2808 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGraziano, F., Valsecchi, M. G. \u0026amp; Rebora, P. Sampling strategies to evaluate the prognostic value of a new biomarker on a time-to-event end-point. \u003cem\u003eBMC Med. Res. Methodol.\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e, 93 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKent, P., Cancelliere, C., Boyle, E., Cassidy, J. D. \u0026amp; Kongsted A. A conceptual framework for prognostic research. \u003cem\u003eBMC Med. Res. Methodol.\u003c/em\u003e \u003cb\u003e20\u003c/b\u003e, 172 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins, G. S. et al. TRIPOD\u0026thinsp;+\u0026thinsp;AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. \u003cem\u003eBMJ\u003c/em\u003e \u003cb\u003e385\u003c/b\u003e, e078378 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTatyana Mollayeva, D., Gi, E. M. E. E. \u0026amp; F.-T. D. G. S. H. B. L. M. P. T. T. S. D. C. Data Templates in A PROGRESS-Driven Approach to Cognitive Outcomes after Traumatic Brain Injury: Advancing Equity, Diversity, and Inclusion through Knowledge Synthesis and Mobilization. \u003cem\u003eOSF\u003c/em\u003e (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilkinson, M. D. et al. The FAIR Guiding Principles for scientific data management and stewardship. \u003cem\u003eSci. Data\u003c/em\u003e. \u003cb\u003e3\u003c/b\u003e, 160018 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSant\u0026rsquo;Ana, T. T. et al. A PROGRESS-driven approach to cognitive outcomes after traumatic brain injury: A study protocol for advancing equity, diversity, and inclusion through knowledge synthesis and mobilization. \u003cem\u003ePLoS One\u003c/em\u003e. \u003cb\u003e19\u003c/b\u003e, e0307418 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMollayeva, T., Pacheco, N., D\u0026rsquo;Souza, A. \u0026amp; Colantonio, A. The course and prognostic factors of cognitive status after central nervous system trauma: a systematic review protocol. \u003cem\u003eBMJ Open.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, e017165 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMollayeva, T. et al. C. (). A systematic review of brain health outcomes after traumatic brain injury through a PROGRESS-Plus lens. \u003cem\u003ePROSPERO\u003c/em\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVo\u0026szlig;, H. et al. HarmonizR enables data harmonization across independent proteomic datasets with appropriate handling of missing values. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 3523 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAdhikari, K. et al. Data Harmonization and Data Pooling from Cohort Studies: A Practical Approach for Data Management. \u003cem\u003eInt J. Popul. Data Sci\u003c/em\u003e \u003cb\u003e6\u003c/b\u003e, (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGill, I. S. et al. The DataHarmonizer: a tool for faster data harmonization, validation, aggregation and analysis of pathogen genomics contextual information. \u003cem\u003eMicrob Genom\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFortier, I., Doiron, D., Burton, P. \u0026amp; Raina, P. Invited Commentary: Consolidating Data Harmonization\u0026ndash;How to Obtain Quality and Applicability? \u003cem\u003eAm. J. Epidemiol.\u003c/em\u003e \u003cb\u003e174\u003c/b\u003e, 261\u0026ndash;264 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoyden, J. \u0026amp; Walnicki, D. \u003cem\u003eLeveraging the Power of Longitudinal Data: Insights on Data Harmonisation and Linkage from Young Lives\u003c/em\u003e. (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL\u0026ouml;w, N., Hesser, J. \u0026amp; Blessing, M. Multiple retrieval case-based reasoning for incomplete datasets. \u003cem\u003eJ. Biomed. Inf.\u003c/em\u003e \u003cb\u003e92\u003c/b\u003e, 103127 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMainzer, R. M. et al. Gaps in the usage and reporting of multiple imputation for incomplete data: findings from a scoping review of observational studies addressing causal questions. \u003cem\u003eBMC Med. Res. Methodol.\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e, 193 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, K. J., Carlin, J. B., Simpson, J. A. \u0026amp; Moreno-Betancur, M. Assumptions and analysis planning in studies with missing data in multiple variables: moving beyond the MCAR/MAR/MNAR classification. \u003cem\u003eInt. J. Epidemiol.\u003c/em\u003e \u003cb\u003e52\u003c/b\u003e, 1268\u0026ndash;1275 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg, S. M. et al. From local explanations to global understanding with explainable AI for trees. \u003cem\u003eNat. Mach. Intell.\u003c/em\u003e \u003cb\u003e2\u003c/b\u003e, 56\u0026ndash;67 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu, J. \u0026amp; Szymczak, S. A review on longitudinal data analysis with random forest. \u003cem\u003eBrief Bioinform\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWallace, M. L. et al. Use and misuse of random forest variable importance metrics in medicine: demonstrations through incident stroke prediction. \u003cem\u003eBMC Med. Res. Methodol.\u003c/em\u003e \u003cb\u003e23\u003c/b\u003e, 144 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, Y. W., Choi, J. W. \u0026amp; Shin, E. H. Machine learning model for predicting malaria using clinical information. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cb\u003e129\u003c/b\u003e, 104151 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAustin, P. C., Harrell, F. E., Lee, D. S. \u0026amp; Steyerberg, E. W. Empirical analyses and simulations showed that different machine and statistical learning methods had differing performance for predicting blood pressure. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 9312 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, H. S., Kim, J. H., Son, J., Park, H. \u0026amp; Choi, J. Machine learning models for predicting early hemorrhage progression in traumatic brain injury. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 11690 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobeson, S. M. \u0026amp; Willmott, C. J. Decomposition of the mean absolute error (MAE) into systematic and unsystematic components. \u003cem\u003ePLoS One\u003c/em\u003e. \u003cb\u003e18\u003c/b\u003e, e0279774 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang, S., Liang, Y., Shi, S., Wu, C. \u0026amp; Shi, Z. Improving predictions and understanding of primary and ultimate biodegradation rates with machine learning models. \u003cem\u003eSci. Total Environ.\u003c/em\u003e \u003cb\u003e904\u003c/b\u003e, 166623 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePedregosa, F. Machine learning in python - extra trees regressor. \u003cem\u003eScikit Learn\u003c/em\u003e (2011). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://scikit-learn.org/stable/modules/generated/sklearn.ensemble.ExtraTreesRegressor.html\u003c/span\u003e\u003cspan address=\"https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.ExtraTreesRegressor.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePedregosa, F. Machine learning in python - gradient boosting regressor. \u003cem\u003eScikit Learn\u003c/em\u003e (2011). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://scikit-learn.org/stable/modules/generated/sklearn.ensemble.GradientBoostingRegressor.html\u003c/span\u003e\u003cspan address=\"https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.GradientBoostingRegressor.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artificial intelligence, Common data elements, Data harmonization, Evidence gap, Information management, Social determinants of health","lastPublishedDoi":"10.21203/rs.3.rs-8585532/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8585532/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Scientific research on social parameters for prognosis after traumatic brain injury (TBI) is evolving, yet results remain heterogeneous, and predictors for risk stratification are lacking. To understand how social parameters are linked to cognitive outcomes after TBI, we applied machine learning (ML) algorithms using data from 30 published studies including 2,364 participants with TBI (72% male; 55% mild, 45% moderate-severe injury). We extracted and harmonised longitudinal data following the PROGRESS-Plus framework, and used the data as predictors of rate of change in cognition post-TBI. We developed random forest, gradient boosting (GB), and extreme GB predictive models, accounting for time from injury to baseline assessment, time between assessments, and country-level structural indicators. We used mean absolute error and root mean squared error to evaluate model performance, and Shapley Additive Explanations analysis for explanatory model predictions. Results highlighted time interval, country-level structural indicators, age, and variation in education as key predictors for rate of change for both injury severities. Sensitivity analyses for predicting rate of change in executive function and learning and memory confirmed the robustness of the results. Our work contributes to novel ML research for understanding prognosis and advancing precision in predicting cognitive outcomes after TBI.","manuscriptTitle":"Explainable machine learning of PROGRESS-Plus social factors predicts cognitive trajectories after traumatic brain injury","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-28 01:26:57","doi":"10.21203/rs.3.rs-8585532/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"906dbb1d-577a-4310-be46-c95e04a6172b","owner":[],"postedDate":"January 28th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":61534812,"name":"Health sciences/Medical research"},{"id":61534813,"name":"Health sciences/Neurology"},{"id":61534814,"name":"Biological sciences/Neuroscience"}],"tags":[],"updatedAt":"2026-04-07T16:14:01+00:00","versionOfRecord":{"articleIdentity":"rs-8585532","link":"https://doi.org/10.1038/s41598-026-44818-5","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-03-30 15:58:43","publishedOnDateReadable":"March 30th, 2026"},"versionCreatedAt":"2026-01-28 01:26:57","video":"","vorDoi":"10.1038/s41598-026-44818-5","vorDoiUrl":"https://doi.org/10.1038/s41598-026-44818-5","workflowStages":[]},"version":"v1","identity":"rs-8585532","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8585532","identity":"rs-8585532","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}