Constructing multicancer risk cohorts using national data from medical helplines and secondary care

Constructing multicancer risk cohorts using national data from medical helplines and secondary care | 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 Constructing multicancer risk cohorts using national data from medical helplines and secondary care Hadi Modarres, Dimitris Pipinis, Divya Balasubramanian, Rupert Chaplin, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5701032/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 27 Aug, 2025 Read the published version in npj Digital Medicine → Version 1 posted 11 You are reading this latest preprint version Abstract Identification of cohorts at higher risk of cancer can enable earlier diagnosis of the disease, which significantly improves patient outcomes. In this study, we use machine learning to predict cancer diagnosis in the next year. We select nine cancer sites with high incidence of late-stage diagnosis or worsening survival rates, and where there are currently no national screening programmes. We use National Health Service (NHS) data from medical helplines (NHS 111) and secondary care appointments from all hospitals in England. We show that features based on information captured in NHS 111 calls are among the most influential in driving predictions of a future cancer diagnosis. Our predictive models exhibit good discrimination (AUC – 0.78 – SD 0.04), ranging from 0.69 (ovarian cancer) to 0.83 (oesophageal cancer). While our predictive modelling provides patient level risk predictions, our emphasis is on constructing cohorts of patients who may be at risk of cancer rather than individual risk scores. We present an approach of constructing cohorts at higher risk of cancer based on feature importance and considering possible bias in model results. These outputs can be used to develop highly targeted case finding services, which could help increase earlier detection rates and reduce health disparities. This approach is flexible and can be tailored based on the group the intervention targets (i.e. symptomatic or asymptomatic patients) and the data available to those charged with administering the intervention. Physical sciences/Mathematics and computing/Statistics Biological sciences/Cancer Health sciences/Health care Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Improving specificity in identifying cohorts at higher risk of developing cancer could increase rates of early diagnosis and allow more focussed interventions to be delivered. However, diagnosing people early is complicated as early-stage symptoms can be harder to definitively attribute to cancer pathology. This means a very large number of individuals would need to be tested to detect a relatively small number of cancers, rendering using such symptoms impractical as a basis for symptomatic case finding or population level screening programmes. To achieve more accurate cancer incidence prediction, the last decade has seen a proliferation of machine learning models trained with unprecedented access to large datasets and computing power. Previous research, in that vein, has typically either i) used routinely collected data, from either secondary or primary care, 1 – 10 or ii) used imaging and/or biomarker/genetic data, which are limited to small segments of the population (e.g. those with specific comorbidities). 11 – 15 Previous research has provided a wealth of findings highlighting the promise of using machine learning for cancer risk prediction. Both these approaches are not however without their limitations when it comes to early identification of cohorts at higher risk of cancer among the general population. This is because secondary care data may only capture cancer specific events that are picked up quite late in the patient pathway, which could result in worse outcomes. On the other hand, focusing only on primary care data may miss the useful information included in secondary care data, which has shown promise in recent research. 16 Moreover, the use of data that are available for only a small segment of the population is unsuitable for identifying high risk cohorts at the population level. To optimise identification of high-risk cohorts, a combination of both primary and secondary care data, at the population level, would be preferable. This approach could capture both symptoms and lifestyle factors, as well as detailed comorbidities, which have previously been shown to be useful signals for predicting future cancer incidence. In the absence of access to national level primary care data, we decided to use National Health Service (NHS) data from medical helplines, specifically, NHS 111 calls data. NHS 111 lines is an alternative source of healthcare advice and information if access to a General Practitioner (GP) is not possible. This dataset records symptoms individuals were concerned about and could provide early insights of undiagnosed cancer, signals which we would miss if relying solely on secondary care data. Our research progresses the field in this direction by matching secondary care data - capturing important features including pre-existing comorbidities, frequency of hospital appointments and demographics - to NHS 111 calls data. This is the first time that data from medical helplines have been used to predict future cancer incidence. Data from NHS 111 calls have the additional advantage that are not “clinician-initiated data” in the sense that “they do not reflect data created through specific actions (or inactions) or insights of the clinician”. This is significant because recent research has made a compelling case that predictions based on “clinician-initiated data” may have limited added value, compared to “what the average clinician would decide for the average similar patient”. 17 Alongside the rich data from secondary care, we construct detailed patient pathways covering six years of patient history. By doing so, we capture comorbidities and frequency of interactions with the healthcare system. There is some evidence that frequency of interactions in secondary care may reflect the number of missed appointments in primary care, arguably a reflection of patient behaviour, and, as such, including frequency of secondary care interactions could help capture relevant patient behaviour towards their health. 18 We focus on nine cancers (bladder, head and neck, kidney, lymphoma, myeloma, oesophageal, ovarian, pancreatic, and stomach), which are associated with a high proportion of late-stage diagnoses (stage III and IV) or worsening survival rates in England, and don’t currently have screening programmes. We focus on predicting the risk of first cancer diagnosis. Building on model results and making use of feature importance, we successfully develop an approach for constructing higher risk cohorts of varying size while minimising the possible bias that may come from the relatively small numbers of patients for certain demographics. Our approach complements a more standard approach of identifying cohorts at higher risk based on individual risk predictions. It affords greater insights by constructing higher-risk cohorts based on feature importance, the data available to those charged with administering the intervention, and the type of intervention (e.g. whether it targets symptomatic or asymptomatic patients). We illustrate this approach, by applying it to bladder cancer, a priority cancer site. Our study also provides the first implementation of multi-cancer prediction modelling using population level data in England. Recent work has made the case for the utility of multi-cancer predictive modelling in the context of new liquid biopsy tests, which are currently under development and evaluation. 18 The results presented here could provide further evidence of the possibilities for multi-cancer prediction afforded by national level health data collections. Results Population description Our dataset includes 23.6 million patient histories of individuals between 40 and 74 years old in England (see methods for full details on dataset construction). This age cohort is selected based on the relatively higher incidence of cancer (compared to younger cohorts), and the fact that diagnostics and treatment are less likely to pose complications (e.g. due to frailty), compared to older cohorts. In order to focus on the first cancer diagnosis, we exclude all those with a previous cancer diagnosis from the study population. Our patient histories cover six years between 2016–2022. We use the first 5 years of the patient histories to predict cancer diagnosis in year 6. A stylised version of a patient pathway diagnosed with cancer in year 6 is presented in Fig. 1 . An individual’s patient history includes not only demographic information, but also comorbidities diagnosed during the 5-year period (between 2016 and 2021), as well as information on symptoms reported to NHS 111 lines. Data on mortality allows us to monitor who may have passed away during the prediction period (September 2021-August 2022) for reasons other than cancer and exclude those individuals from our data. We also exclude individuals with a previous cancer diagnosis. More details on dataset construction are included in the methods section. Model prediction We trained several classification models to predict the probability of being diagnosed with cancer in the coming year (September 2021-August 2022). We selected the XGBoost model as our preferred specification based on comparisons in terms of performance with the other classifiers. Given the very sharp class imbalance between cancer and non-cancer cases, we use under sampling in our training datasets to ensure an equal number of cancer and non-cancer cases. We predict the risk of cancer diagnosis for the nine cancer sites selected for the reasons discussed earlier. For each of these models, we report several performance metrics for all cancer specific models (see Table 1 ). Table 1 Model performance on the test dataset across different cancer sites. Cancer site AUC Accuracy Specificity Sensitivity Bladder 0.818 0.731 0.731 0.755 Head and neck 0.756 0.669 0.669 0.709 Kidney 0.785 0.707 0.707 0.715 Lymphoma 0.747 0.663 0.663 0.705 Myeloma 0.780 0.706 0.706 0.697 Oesophageal 0.831 0.707 0.707 0.800 Ovarian 0.693 0.625 0.625 0.658 Pancreatic 0.809 0.671 0.671 0.796 Stomach 0.792 0.682 0.682 0.752 Important features In Table 2 , we present some descriptive statistics focusing on the comparisons between those diagnosed with bladder cancer during the predictive window and those who were not. Table 2 Descriptive statistics bladder cancer diagnosis vs no bladder cancer diagnosis in the predictive window (year 6). Bladder cancer diagnosis No bladder cancer diagnosis Count 7130 23620663 Age 64.7 (7.8) 55.3 (9.7) Gender (male/female) 75.9%/24.1% 50.8%/49.2% Ethnicity (White British/Other) 85.9%/14.1% 71.6%/28.4% Residing in 5 most deprived IMD deciles (%) 48.9% 48.2% NHS 111 calls reporting cancer related symptoms (during year 5) 0.17 (0.58) 0.05 (0.4) A&E attendances (during year 5) 0.54 (1.18) 0.26 (0.85) IMD stands for Index of Multiple Deprivation. Standard deviation in parenthesis. As expected, there are noticeable differences in terms of age and gender between those diagnosed with bladder cancer in year 6 and those who were not. Cancer cases are predominantly male and older, reflecting the well-established link between age and cancer incidence, as well as the fact that bladder cancer is more frequent among males. Beyond demographics, the number of 111 calls reporting cancer related symptoms (see Table 6 for full list), as well as the number of A&E attendances, are higher on average for those diagnosed with bladder cancer in year 6 compared to those who are not. Both the number of calls to NHS 111 lines reporting cancer related symptoms, as well as number of A&E attendances, will end up being among the most useful features for model prediction, as we will see later. In order to improve model accuracy and to inform our work on constructing higher risk cohorts, we select the most important features using two metrics, gain and Shapley (SHAP) values. We demonstrate our approach for cohort construction by using bladder cancer as our test case. In Fig. 2 a, we show the SHAP values for the top 20 features (our models include more than 800 features in total) – ordered based on the gain metric (the average gain across all splits where the feature is used) for the XGBoost model (Fig. 2 b). The red colour indicates higher values for the selected feature, and a positive SHAP value means an increase in the risk of cancer. For example, higher age (red colour) has overwhelmingly positive SHAP values, which means that higher age is predicting higher risk of bladder cancer in the next year. By comparison, we show the mean absolute SHAP value for these features (Fig. 2 c). While there are slight differences in the ordering of the most informative features, 17 out of the top 20 features based on average model gain are also amongst the top 20 features as determined by mean absolute SHAP value. We observe that beyond age and gender, several comorbidities appear as relevant predictors of a bladder cancer diagnosis in ways that are consistent with expectations based on the medical literature. For example, the presence of chronic obstructive pulmonary disease (COPD) and urinary infections is associated with the incidence of bladder cancer in previous research. 20 , 21 In addition, several features drawn from the NHS 111 calls dataset appear to be good predictors of bladder cancer incidence. For example, higher number of calls to NHS 111 lines reporting cancer related symptoms is one of the features with the highest gain metric value (just below demographics and long-term condition status). In addition, we also see that features capturing specific symptoms that are plausibly related to undiagnosed bladder cancer are also relevant and have the expected direction of effect. Specifically, higher number of calls to NHS 111 lines reporting “pain and frequency of passing urine” or “blood in urine” (during the last year) are both relevant predictors of risk of being diagnosed with bladder cancer in the next year. To more comprehensively explore the importance of NHS 111 calls as predictors of risk of future cancer diagnosis, we replicated the analysis based on the gain metric for all other priority cancers beyond bladder. In all cases, features based on NHS 111 calls were among the most influential in predicting future cancer diagnosis. In Table 3 , we report the rank of features, created from NHS 111 calls data, in terms of feature importance based on the gain metric. We do this for all cancer sites included in Table 1 . Table 3 Feature importance rank based on the gain metric for features based on NHS 111 calls. Cancer site Feature name (rank) Bladder Number of 111 calls in last year (4th ) Number of 111 calls relating to pain/frequency of passing urine in last year (11th ) Number of 111 calls relating to blood in urine in last year (13th ) Head and neck Number of 111 calls in last year (17th ) Kidney Number of 111 calls related to cancer symptoms in last year (2nd ) Number of 111 calls in last year (9th ) Lymphoma Number of 111 calls in last year (5th ) Number of 111 calls related to cancer symptoms in last year (10th ) Number of 111 calls in previous 5 years(13th ) Myeloma Number of 111 calls in last year (5th ) Number of 111 calls in previous 5 years (15th ) Number of 111 calls related to cancer symptoms in last year (17th ) Oesophageal Number of 111 calls in last year (10th ) Number of 111 calls related to cancer symptoms in last year (11th ), Ovarian Number of 111 calls in last year (1st ) Number of 111 calls related to cancer symptoms in last year (5th ) Number of 111 calls relating to chest and upper back pain in last year (9th ) Pancreatic Number of 111 calls in last year (5th ) Number of 111 calls related to cancer symptoms in last year (6th ) Number of 111 calls relating to abdominal pain in last year (12th ) Stomach Number of 111 calls in last year (5th ) Number of 111 calls related to cancer symptoms in last year (7th ) Table 3 highlights, that features based on information captured in NHS 111 calls are among the top 20 features, based on the gain metric, and often among the top 5 or 10 for all cancer sites we explored in this study. Constructing higher-risk cohorts The primary goal for the analysis is to use the model results to construct high risk cohorts, for a cancer diagnosis within the next year, which can then be used to inform case finding and appropriate interventions to support earlier diagnosis and improve survival. We discuss two possible approaches to achieving this. In the first method (Method A), we use the model risk probability at the individual patient level to create cohorts. Different sized cohorts can be constructed by varying the threshold for inclusion in the high-risk group. The second method (Method B) identifies cohorts with defined characteristics based on decision rules, utilising the most informative features from the trained model. Method A One approach to constructing higher risk cohorts is to consider capacity based on the requirements of a specific intervention/screening programme, and then selecting the appropriate risk threshold which would lead to the desired cohort size. The risk thresholds are applied to the individual level predictions of the model. Based on different risk thresholds, high-risk cohorts of varying sizes can be constructed. The lower the risk threshold, the larger the size of the cohort, but the lower the potential incidence of cancer within the cohort. An illustration of this method is shown in Fig. 3 . We define a lift value as the ratio of the cancer incidence within the cohort to the baseline cancer incidence. The baseline cancer incidence in our case, refers to those aged between 40–74 with no previous cancer diagnoses. Based on different risk thresholds, one could construct a lift curve which plots the lift value against the size of the cohort on the x-axis. An example based on the model on bladder cancer is provided in Fig. 4 . The lift curve exhibits the expected shape where the lift value declines as we increase the size of the cohort. It also highlights the potential trade-off between high incidence and the total number of cancers correctly predicted. As the cohort size is increased, individuals with lower risk scores are included in the cohort, reducing the incidence and hence the lift value. Typically, the smaller the selected cohort of the population, the higher the lift value, as these are the individuals with the highest risk scores from the model. For example, in the top left of the figure, considering a cohort size of the highest risk 125,000 individuals (based on model probability risk score), the lift value of the model trained on all variables is 16.6 (representing a cancer rate of 1 in 202 in the cohort). For a model trained on only demographic variables, the lift value for an equivalent cohort size is 9.4 (representing a cancer rate of 1 in 357). These values represent the potential order of magnitude improvement in cancer incidence in identifying high-risk cohorts using a risk score approach compared to chance selection from the population (cancer rate of 1 in 3355). If the cohort size is increased to 1 million individuals at highest risk, the lift value reduces to 7.2 for the model with all variables and 5.7 for the model with demographic variables only. These results also demonstrate the importance of including features from calls to NHS 111 lines, as well as comorbidities, to the model, as the resulting model risk scores can identify the highest risk individuals more accurately. As the cohort size increases, the difference in lift value between the two reduces, suggesting that both models capture the background demographic risk factors. While more standard in terms of approach, Method A has several limitations from an operational perspective. First, Method A does not allow one to filter on features that may be most useful from a practical perspective. For example, depending on the type of intervention, one may want to focus on symptomatic patients and therefore select cohorts based on specific symptoms. Second, those charged with administering the interventions may not always have access to individual level predictions and instead would have to rely on flags drawing from specific features that are included in the data available to them. For example, eligibility for targeted lung health checks in England relies on age and smoking status. This method also relies on applying thresholds to individual patient level predictions. While steps can be taken to explain the model (e.g. through feature importance and other techniques), ultimately the high-risk groups are a heterogeneous cohort. In Method B, we demonstrate how clearly defined cohorts can be created based on specific feature combinations. A further limitation of this method is that the model may be biased towards predicting higher risk for certain demographic groups, making the high-risk cohort non representative of the actual incidence of bladder cancer in the population. As shown in the SHAP feature importance results (Fig. 2 a), higher age, male, and white ethnicity all tend to increase the model risk score. This does correspond with higher incidence of bladder cancer in this group, however, given the low counts of bladder cancer among other demographic groups, it is difficult to ensure fair representation of all strata when constructing high-risk cohorts using this method, even when steps are taken to balance the training dataset. In the sub-group cohorts of Method B, we show how we sought to address this. Method B An alternative method that overcomes the above constraints is shown in Fig. 5 and outlined as follows. First, we select relevant features based on feature importance and data availability. The most important features, which were in the top 20 of model gain and SHAP value, were selected. SHAP was also used to identify the direction of the feature. Features which had a positive impact on the model output (i.e. which tended to increase the risk if the feature was present) were selected. We then filter the population based on those features and examine the predicted cancer incidence. As is the case with Method A, we can then compare incidence of cancer in this curated cohort compared to the baseline incidence in the entire population. This method has been applied to the whole population, and to sub-groups of demographic strata, demonstrating how the approach can be used for targeted screening. Population wide cohorts The pair of features which would yield the highest incidence cohorts (on the validation data) of varying size (at least 10,000 to at least 250000) were identified. Subsequently, the selected pair of features was applied to the test data to evaluate the expected bladder cancer incidence in the whole population. In Table 4 , we show some examples of such curated cohorts based on combinations of just two features among those that the model considers as high importance for predicting cancer incidence in the next year. The cohort with highest cancer incidence, and a size of at least 10,000, is constructed based on interactions with the 111-call service and includes a specific bladder cancer related symptom of blood in urine. The cancer incidence within this cohort is 41 times higher than the overall incidence in the analysis population, with a cancer incidence of 1 in 82, compared to 1 in 3355 in the study population. Larger cohorts of high-risk patients are constructed with flags relating to comorbidities of the genitourinary system and other diseases of the urinary system. Applying these flags to the population results in a cohort size of approximately 100,000 individuals, with a cancer rate 6 times higher than in the overall study population. An example of a larger cohort of ~ 290,000 individuals would be constructed by applying the filter of patients having at least one long-term condition, and a diagnosis relating to symptoms and signs involving the genitourinary system in the last 5 years. This results in a cohort with a lift value of 4.5. Table 4 Example high-risk cohorts of varying sizes, applied to the whole population. Clearly defined rules, based on the most informative model features, are used to construct the cohorts of varying sizes. Higher cancer incidence is found in smaller cohorts. Feature combination Population size Incidence in cohort (%) Lift value At least one call reporting cancer related symptoms in last year AND at least one call reporting blood in urine in last year 16,700 1.2% 41 Diagnosis of “Symptoms and signs involving the genitourinary system” (ICD10 R30-R39) in the last 5 years AND Diagnosis of “Other diseases of the urinary system” (ICD10 N30-N39) in last 5 years 98,300 0.18% 6 Has a long-term condition AND Diagnosis of “Symptoms and signs involving the genitourinary system” (ICD10 R30-R39) in the last 5 years 290,000 0.14% 4.5 Sub-group cohorts The population was segmented into demographic groups to investigate if different sets of features can create higher risk cohorts across demographic strata. This was also to address one of the limitations of method A, namely how we can ensure equality of opportunity if model predictions may be biased when there is insufficient training data from all demographic groups. The segmentation was based on gender (male/female) and broad ethnicity (White/Non-white), resulting in four groups. Due to the low incidence of bladder cancer, more granular segmentation would have resulted in very small sample sizes. For each population segment, the same methodology as described above was applied, with the cohort with the highest incidence of cancer cases being identified. These decision rules were then applied to the test dataset to evaluate the efficacy of the cohort. The lift value was calculated based on the incidence of cancer for each stratum. The results are shown in Table 5 . Table 5 Example high-risk cohorts applied to demographic strata. For each stratum, the feature combination which resulted in the highest cancer incidence in the cohort (of minimum size 5000) is shown. Demographic strata Feature combination Cohort size Incidence in cohort (%) Lift value Female – non white ethnicity At least 1 A&E attendance in the last year AND diagnosis of a long term condition 195000 0.01% 2.9 Female – white ethnicity At least one call reporting cancer related symptoms in last year AND at least one call reporting blood in urine in last year 6800 0.8% 47.5 Male – non white ethnicity At least 1 A&E attendance in the last year AND diagnosis of COPD 7000 0.07% 4.9 Male –white ethnicity At least one call reporting cancer related symptoms in last year AND at least one call reporting blood in urine in last year 7600 1.9% 36 For the white ethnicity group, features related to 111 calls are particularly effective in identifying high-risk groups. The specific nature of the symptom information (blood in urine) can result in small cohorts with lift values of 47.5 for white females, and 36 for white males. In contrast, for the non-white ethnic group, more general health factors (e.g. A&E attendance) and comorbidities (e.g. COPD) result in the highest risk groups. These cohorts are still significantly higher in cancer incidence compared to baseline rates for these populations, as shown by the lift values of 2.9 for females, and 4.9 for males. However, they are also significantly lower than the lift values obtained for the white ethnic group. This likely reflects health inequalities in the utilisation of services such as 111 calls. Discussion This study makes several contributions to the burgeoning literature that seeks to use machine learning to develop useful predictive models for cancer incidence. First, our results demonstrate that information included in medical helplines such as NHS 111 calls contains useful signals predicting a future cancer diagnosis. Our results, show that without exception for the nine cancer types we examined, features based on NHS 111 calls are among the most significant in terms of importance for predicting a future cancer diagnosis. While data quality and coverage are high when it comes to reported symptoms in NHS 111 calls, this dataset is not as comprehensive as primary care datasets. Future work should look to leverage those datasets alongside the information included in secondary care and NHS 111 calls to create a more complete patient history. The second contribution of the study is to describe a practical method of constructing higher risk cohorts that could be tailored based on data availability, type of intervention, and desired levels of accuracy. We showcase this approach drawing from the model predictions for bladder cancer incidence. Beyond its greater flexibility, our approach also mitigates the potential for bias due to the underrepresentation of certain demographic groups in the data. Finally, ours is the first study employing multi-cancer prediction modelling using population level data from England. Our models exhibit good performance for most cancer types. These results further strengthen the case for using routinely collected national health data to risk stratify the population based on risk for future cancer incidence. There are numerous potential practical applications of this analysis. Data could be used to inform case finding services for the high-risk cohorts identified. Such case finding services may be used to identify populations at a higher risk of developing cancer, who would benefit from ongoing surveillance, as well as individuals who may warrant an urgent diagnostic test for cancer. Populations with red flag symptoms of cancer, who meet referral thresholds indicated in NG12 (NICE guidelines on suspected cancer), could be triaged directly into urgent suspected cancer pathways. 22 Cohort characteristics could also be used to inform and better target opportunistic cancer checks, as well as local public awareness campaigns, reflecting symptom combinations and/or geographies with increased risk. Our study is not without its limitations. For one, coverage of early reported symptoms of underlying disease is not as complete as it would have been if we had been able to use data from primary care. Second, predicting future cancer incidence does not differentiate based on stage of cancer diagnosis. It could be argued that predicting cohorts at risk of presenting at a late stage of cancer would be more valuable in terms of improving early diagnosis, as it could allow better targeting of those interventions to such populations. This information is available in the cancer registry database in England but was not available to the authors at the time of this research. Finally, our study does not include any information on lifestyle factors, which almost certainly play an important role in affecting the baseline risk of future cancer incidence. This is a limitation that is common in much of the previous work relying on secondary care data and this study is not going beyond previous research in that regard. Methods Datasets Our predictive models are trained on a dataset that captures an individual’s previous interactions with the healthcare system, their comorbidities, as well as a rich set of socio-demographic information. To create these patient histories, we combine several large datasets including the National Bridges to Health Segmentation Dataset, Secondary Use Services (SUS) data, Emergency Care Data Set (ECDS), NHS 111 calls data, as well as ONS mortality data. 23 , 24 A brief description of these datasets is provided below. The National Bridges to Health Segmentation Dataset (B2H), which draws on a large number of datasets, provides information on long-term conditions for all patients, over 60 million, registered with General Practices in England. 25 In addition, we use the information included in B2H for socio-demographic characteristics (e.g. including race, age, sex, deprivation, household type) that could affect the risk of a future cancer diagnosis. To complement the information included in B2H, we draw on data from the SUS and ECDS datasets, which include information on all outpatient, inpatient and emergency attendances in hospitals in England. This allows us to capture information on the number of previous inpatients/outpatients and emergency attendances. Frequency of interactions with the health system could reflect attitudes towards one’s health, beyond capturing underlying healthcare needs. For this reason, we construct several features which capture separately the number of previous hospital admissions, outpatient appointments, and emergency attendances within different time periods in the past (e.g. last year, last 5 years). The SUS/ECDS datasets also allow us to capture detailed comorbidities as diagnosed in secondary care. We use ICD 10 codes covering 263 groups of comorbidities. A key dataset used in the analysis is the NHS 111 calls dataset. This dataset covers all calls made to NHS 111 lines between 2018–2023. It includes information on the symptoms the caller reported, as well as the date of the call. With the help of clinicians, we identified symptoms which may be related to cancer and constructed the relevant features capturing both the frequency of any cancer related symptoms reported, as well as the frequency for specific symptoms (e.g. blood in urine). Table 6 with the cancer related symptoms is below. Table 6 Cancer related symptoms reported in NHS 111 calls. Cancer related symptoms reported in NHS 111 calls Abdominal flank groin or back pain or swelling Abdominal pain pregnant over 20 weeks Abdominal pain rectal bleeding pregnant over 20 weeks Abdominal pain Blood in urine Breast lump pregnant Breast lump Breathing problems breathlessness or wheeze pregnant Breathing problems breathlessness or wheeze Chest and upper back pain Constipation Cough Coughing up blood Diarrhoea Difficulty passing urine Easy or unexplained bruising ED Triage chest pain Face neck pain or swelling Fever Genital problems Itch Mouth ulcers Pain and/or frequency passing urine Rectal bleeding Rectal pain swelling lump or itch Skin lumps Skin problems Tiredness fatigue Urinary problems Vaginal bleeding Vaginal discharge Vomiting Vomiting blood Finally, to account for censoring due to death, we use person level data from the Office for National Statistics (ONS) death on mortality to capture those passing away during the period. The various datasets are linked together using the pseudonymised ID that is common across the datasets. This allows us to create patient histories that capture all patient interactions with secondary care in the NHS, as well as any calls to the NHS 111 lines. Alongside any diagnosed comorbidities and sociodemographic information, this dataset provides rich information on which to build our predictive modelling. Feature construction and preprocessing In the section below, we describe in more detail how the features are constructed. The complete variable list of features used for modelling is listed in Supplementary Table 1). Comorbidity For each ICD10 category 3 code block (e.g. A00-A09, A15-A19 … Z80-Z99), we create a flag per patient, to indicate if they received a diagnosis in this category, in the last year or last 5 years. This was done by evaluating all diagnosis fields from SUS Inpatient, outpatient, A&E, and the emergency care dataset (ECDS). Interactions with the healthcare system The number of attendances at A&E, inpatient, and outpatient settings was calculated for each patient in the last year and the last 5 years from the cut-off date. The number of calls to 111 in the last year, and the number of calls with potentially cancer related symptoms was calculated. Socio-demographic We use one-hot encoding for categorical variables (Ethnicity, Index of multiple deprivation, Integrated Care Board, Acorn household type). The latter variables segment households into 6 categories (and 62 types) capturing financial circumstances, benefit receipt, health, wellbeing, and leisure and shopping behaviours. 26 In addition we include age, as well as an indicator variable capturing whether the individual is residing in a care home. For modelling purposes, we also impose several exclusions which we discuss below. Exclusions and missing data The focus of this analysis is individuals who are aged between 40 and 74 at the cut-off date between the observation and prediction window. We therefore exclude younger cohorts who are likely to have a much lower risk of developing cancer, as well as older individuals. We then impose the following restrictions. We exclude those with a previous cancer diagnosis, as our focus here is on first diagnosis of cancer. We also exclude those who passed away before the cut-off date as well as those who passed away for any other reason than cancer after the cut-off date. We finally exclude the small number of individuals (1824) with missing information on gender. Where data was missing for categorical variables, the null value was replaced with an ‘unknown’ string value. For those with missing data on ethnicity, we create an “unknown” flag and include this in the analysis. Machine learning analysis The dataset was split into train (60%), validation (20%), and test (20%) datasets through a random split. We performed a series of statistical tests to examine whether there were still systematic differences between the datasets in terms of demographics. No differences were observed between the datasets (details are included in Supplementary Table 3). The size of the datasets is shown in Supplementary Table 4. For training models, the train dataset was randomly under sampled to ensure an equal number of cancer and non-cancer cases. This was done to avoid the issues stemming from large class imbalance due to the very small incidence of cancer in the data. We trained four machine learning models: logistic regression, support vector machine with linear kernel, random forest, and XGBoost. Hyperparameter optimisation was performed by optimising the receiver operating curve area under the curve, using the train and validation datasets with the hyperopt package. The hyperparameters and the ranges for optimisation are provided in the Supplementary Table 2. We report model performance using the test dataset. XGBoost appears to perform better compared to the alternatives, a fact consistent with its reputation in terms of performance when it comes to tabular data. 27 Analysis was performed using python 3.10 on a spark cluster (3.5.0). Versions of the key packages used in the analysis are described in Supplementary Table 5. Training features The XGBoost model was trained with all features (listed in Supplementary Table 1) and also with only demographic and socio-economic variables (age, gender, ethnicity, index of multiple deprivation, geographical variable (Integrated Care Board), care home flag, and acorn household type in order to explore the impact of features relating to 111 calls, comorbidities, and healthcare interactions on the model performance and high-risk cohorts. Feature importance Model feature importance was obtained from the XGBoost model by ranking features by their average gain across all splits the feature is used in. SHAP values were calculated on the validation dataset from trained models. SHAP calculates the contribution of each variable to the model predicted probability output. 28 Cohort construction Method A Individual patient level predictions were obtained on the test dataset. High-risk cohorts were constructed by varying the risk threshold and evaluating the cancer incidence within the cohort. The cohort size (as shown in the lift curve in Fig. 4 ) was obtained by extrapolating to the whole population from the test dataset (which is a random sample comprising 20% of the whole study population). Method B The top 20 most informative features from model gain and SHAP were identified. Features which were present in both lists, and which tended to increase the risk if the feature was present, were selected. Demographic (gender and ethnicity) features were not selected as they were used to segment the population in the sub-group cohorts. Each pair of selected features was used to filter the validation dataset. The size and incidence in the resulting cohort were calculated. For a particular cohort size, the combination of features which resulted in the highest cancer incidence was identified. This pairing of features was then applied to the unseen hold out test dataset to calculate the expected cancer incidence in the wider population. The cohort size in the whole study population was obtained by extrapolating from the test dataset. For sub-groups, the same process as above was applied, with the difference that an additional filtering of the data by demographic strata was also applied. Declarations Ethical Approval Not applicable. Data is collected and used in line with NHS England’s purposes as required under the statutory duties outlined in the NHS Act 2006 and Health and Social Care Act 2012. Data is processed using best practice methodology underpinned by a Data Processing Agreement between NHS England and Outcomes Based Healthcare Ltd (OBH), who produce the Segmentation Dataset on behalf of NHS England. This ensures controlled access by appropriate approved individuals, to anonymised/pseudonymised data held on secure data environments entirely within the NHS England infrastructure. Data is processed for specific purposes only, including operational functions, service evaluation, and service improvement. Where OBH has processed data, this has been agreed and is detailed in a Data Processing Agreement. The data used to produce this analysis has been disseminated to NHS England under Directions issued under Section 254 of the Health and Social Care Act 2012. Author Contribution D.P. and H.M. contributed equally as co-first authors to this work. D.P. wrote the main manuscript text and contributed to the design of the analytical methodology. H.M. led on the design of the analytical methodology, contributed to the writing of the main text/supplementary material and to the analysis. D.B. led on the design of the analytical methodology, contributed to the writing of the main text and to the analysis. S.K. contributed to the analysis and to the writing of the main text/supplementary material. G.T., R.Ch., A.M. contributed to the scoping of the analysis, the design of the analytical methodology and manuscript revisions. R.Ca., E.H.W. contributed to the scoping of the analysis and the writing of the main manuscript. All authors reviewed the manuscript. Acknowledgement The authors would like to thank Robert Scott, and Thomas Henstock for their contributions in the scoping and analysis at the early stages of this study. We would also like to thank Michael Spence and Rajun Phagura for their help at various stages of the analysis. Finally, we are grateful to Anthony Cunliffe GP, Afsana Bhuyia GP, Tina George GP and Amelia Randle GP for sharing their clinical expertise. Data Availability All data used in this study are held internally by NHS England. The data cannot be shared publicly as they contain patient level sensitive information. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. CODE AVAILABILITY Link to data processing notebook: https://github.com/nhsengland/cancer_foundry_data_modelling/ Code for data modelling available upon request from the authors. We plan to publish this code in the near future. References Appelbaum, L., et al., 2021. Development and validation of a pancreatic cancer risk model for the general population using electronic health records: An observational study. European Journal of Cancer, 143, pp.19–30. Wang, Y.H., Nguyen, P.A., Islam, M.M., Li, Y.C. and Yang, H.C., 2019. Development of deep learning algorithm for detection of colorectal cancer in EHR data. In: MEDINFO 2019: Health and Wellbeing e-Networks for All . IOS Press, pp.438–441. Wang, X., et al., 2019. Prediction of the 1-year risk of incident lung cancer: Prospective study using electronic health records from the State of Maine. Journal of Medical Internet Research, 21(5), p.e13260. Hippisley-Cox, J. and Coupland, C., 2015. Development and validation of risk prediction algorithms to estimate future risk of common cancers in men and women: Prospective cohort study. BMJ Open, 5, p.e007825. Malhotra, A., et al., 2021. Can we screen for pancreatic cancer? Identifying a sub-population of patients at high risk of subsequent diagnosis using machine learning techniques applied to primary care data. PLOS ONE, 16(6), p.e0251876. Briggs, E., et al., 2022. Machine learning for risk prediction of oesophago-gastric cancer in primary care: Comparison with existing risk-assessment tools. Cancers , 14, p.5023. Ioannou, G.N., et al. 2018. Development of models estimating the risk of hepatocellular carcinoma after antiviral treatment for hepatitis C. Journal of Hepatology, 69(5), pp.1088–1098. Ioannou, G.N., et al., 2020. Assessment of a deep learning model to predict hepatocellular carcinoma in patients with hepatitis C cirrhosis. JAMA Network Open, 3(9), p.e2015626. Howell, D., et al., 2023. Developing a risk prediction tool for lung cancer in Kent and Medway, England: Cohort study using linked data. BJC Reports , 1, p.16. Zhen, J., et al., 2024. Development and validation of machine learning models for young-onset colorectal cancer risk stratification. npj Precision Oncology , 8, p.239. Steinbuss, G., et al. 2021. Deep learning for the classification of non-Hodgkin lymphoma on histopathological images. Cancers , 13(10), p.2419. Gaur, L., Bhandari, M., Razdan, T., Mallik, S. and Zhao, Z., 2022. Explanation-driven deep learning model for prediction of brain tumour status using MRI image data. Frontiers in Genetics , 13, p.822666. Fontanillas, P., et al., 2021. Disease risk scores for skin cancers. Nature Communications , 12, p.160. Tammemägi, M.C., et al., 2019. Development and validation of a multivariable lung cancer risk prediction model that includes low-dose computed tomography screening results: A secondary analysis of data from the National Lung Screening Trial. JAMA Network Open, 2, p.e190204. Varma, A. et al., 2023. Early prediction of prostate cancer risk in younger men using polygenic risk scores and electronic health records. Cancer Med, 12, pp. 379–386. Placido, D., et al., 2023. A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories. Nature Medicine, 29, pp.1113–1122. Beaulieu-Jones, B.K., et al., 2021. Machine learning for patient risk stratification: Standing on, or looking over, the shoulders of clinicians? npj Digital Medicine, 4, p.62. Williamson, A.E., McQueenie, R., Ellis, D.A., McConnachie, A. and Wilson, P., 2021. ‘Missingness’ in health care: Associations between hospital utilization and missed appointments in general practice. A retrospective cohort study. PLOS ONE, 16(6), p.e0253163. Jung, A.W., Holm, et al. 2024. Multi-cancer risk stratification based on national health data: A retrospective modelling and validation study. The Lancet Digital Health, 6(6), pp.e396-e406. Shephard, E.A., et al. 2012. Clinical features of bladder cancer in primary care. British Journal of General Practice, 62(602), pp.e598-e604. Liao, S., Wang, Y., Zhou, J., Liu, Y., He, S., Zhang, L., Liu, M., Wen, D., Sun, P., Lu, G., Wang, Q., Ouyang, Y. and Song, Y., 2024. Associations between chronic obstructive pulmonary disease and ten common cancers: Novel insights from Mendelian randomization analyses. BMC Cancer , 24(1), p.601. National Institute for Health and Care Excellence. Nice guideline, NG12: Suspected cancer: recognition and referral , last updated 2 October 2023, https://www.nice.org.uk/guidance/ng12 . NHS Digital. Secondary Uses Services (SUS) . https://digital.nhs.uk/services/secondary-uses-service-sus . NHS Digital. Emergency Care Data Set (ECDS) . https://digital.nhs.uk/data-and-information/data-collections-and-data-sets/data-sets/emergency-care-data-set-ecds . Valabhji, J., et al., 2024. Prevalence of multiple long-term conditions (multimorbidity) in England: A whole population study of over 60 million people. Journal of the Royal Society of Medicine, 117(3), pp.104–117. Acorn, 2013. The Household Acorn User Guide. Available at: https://www.caci.co.uk/wp-content/uploads/2021/07/Household_Acorn_UG.pdf [Accessed 16 December 2024]. Borisov, V., et al. 2022. Deep neural networks and tabular data: A survey. IEEE Transactions on Neural Networks and Learning Systems. Lundberg, S.M. and Lee, S.I., 2017. A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems . Curran Associates Inc., pp.4768–4777. Additional Declarations No competing interests reported. Supplementary Files Constructingmulticancerriskcohortssupplementary.docx Cite Share Download PDF Status: Published Journal Publication published 27 Aug, 2025 Read the published version in npj Digital Medicine → Version 1 posted Editorial decision: Revision requested 08 Apr, 2025 Reviews received at journal 22 Mar, 2025 Reviewers agreed at journal 09 Mar, 2025 Reviewers agreed at journal 07 Mar, 2025 Reviews received at journal 26 Feb, 2025 Reviewers agreed at journal 11 Feb, 2025 Reviewers agreed at journal 21 Jan, 2025 Reviewers invited by journal 18 Jan, 2025 Editor assigned by journal 30 Dec, 2024 Submission checks completed at journal 30 Dec, 2024 First submitted to journal 23 Dec, 2024 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-5701032","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":416240222,"identity":"b32491ad-ba62-4dad-8ab6-1e8454b338dd","order_by":0,"name":"Hadi Modarres","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA30lEQVRIiWNgGAWjYBACAyA6AKQZG8DcCoYEUrWcgWo5QEALA1wLYxsRWszZD288wJhjI9s/++zDhz/n3cmTb+A9+PgDHi2WPWkFBxi3pRnPOJdubMy77VmxwQG+ZAO8DjuQYwDUcjix4QwbmzSIsYGBx0wCr5bzbyBa5p9hY//5cw6Q0cBj/gOvlhtQWzYAbWHgbQBad4DHDH+I3XhWcCAR6JeNZ9iYpXmOAfUe5jGWOIPXYcmbP3zcZiM77wwb48cfNUCHtfcYfqjAowUMElB4zISUj4JRMApGwSggCADqBViHB92VrwAAAABJRU5ErkJggg==","orcid":"","institution":"National Health Service England","correspondingAuthor":true,"prefix":"","firstName":"Hadi","middleName":"","lastName":"Modarres","suffix":""},{"id":416240223,"identity":"b8cac09c-df44-495d-899d-1b59b375ce9e","order_by":1,"name":"Dimitris Pipinis","email":"","orcid":"","institution":"National Health Service England","correspondingAuthor":false,"prefix":"","firstName":"Dimitris","middleName":"","lastName":"Pipinis","suffix":""},{"id":416240224,"identity":"5aed231e-b633-4e24-93b2-00b97d0b1565","order_by":2,"name":"Divya Balasubramanian","email":"","orcid":"","institution":"National Health Service England","correspondingAuthor":false,"prefix":"","firstName":"Divya","middleName":"","lastName":"Balasubramanian","suffix":""},{"id":416240225,"identity":"434e22c0-1029-4c28-966b-1341b741764f","order_by":3,"name":"Rupert Chaplin","email":"","orcid":"","institution":"National Health Service England","correspondingAuthor":false,"prefix":"","firstName":"Rupert","middleName":"","lastName":"Chaplin","suffix":""},{"id":416240226,"identity":"478ba7a0-7405-4f4a-8c6e-c9cc7f9e46ae","order_by":4,"name":"Scarlett Kynoch","email":"","orcid":"","institution":"National Health Service England","correspondingAuthor":false,"prefix":"","firstName":"Scarlett","middleName":"","lastName":"Kynoch","suffix":""},{"id":416240227,"identity":"1641b6b3-71eb-4252-978c-953672ceb420","order_by":5,"name":"Achut Manandhar","email":"","orcid":"","institution":"National Health Service England","correspondingAuthor":false,"prefix":"","firstName":"Achut","middleName":"","lastName":"Manandhar","suffix":""},{"id":416240228,"identity":"633c3e7f-2951-4f04-a908-4fbd9314662a","order_by":6,"name":"Gusimran Thandi","email":"","orcid":"","institution":"National Health Service England","correspondingAuthor":false,"prefix":"","firstName":"Gusimran","middleName":"","lastName":"Thandi","suffix":""},{"id":416240229,"identity":"aab2b690-17bd-45d3-b3ab-02fa7538a3fc","order_by":7,"name":"Rebecca Cavilla","email":"","orcid":"","institution":"National Health Service England","correspondingAuthor":false,"prefix":"","firstName":"Rebecca","middleName":"","lastName":"Cavilla","suffix":""},{"id":416240230,"identity":"fe4e28e5-8a88-43c7-9b21-c1c958a71458","order_by":8,"name":"Emma Hirst-Williams","email":"","orcid":"","institution":"National Health Service England","correspondingAuthor":false,"prefix":"","firstName":"Emma","middleName":"","lastName":"Hirst-Williams","suffix":""}],"badges":[],"createdAt":"2024-12-23 16:12:01","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5701032/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5701032/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41746-025-01855-0","type":"published","date":"2025-08-27T15:58:11+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":76833144,"identity":"430a5c11-66b3-4522-abbe-94c5ad2acc77","added_by":"auto","created_at":"2025-02-21 08:48:52","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":84879,"visible":true,"origin":"","legend":"\u003cp\u003ePatient pathway. The figure shows a stylised patient pathway with various types of events recorded during the 5 (history) + 1 (predictive window) years we observe the patients.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-5701032/v1/44d65aa07e7b165792fef4b8.png"},{"id":76834417,"identity":"b02e30e4-54b7-4909-840a-67964a1a79ef","added_by":"auto","created_at":"2025-02-21 09:04:59","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":247368,"visible":true,"origin":"","legend":"\u003cp\u003e(a) SHAP value, shown by order of the top 20 features based on model gain value.\u003cstrong\u003e \u003c/strong\u003eRed (Blue) colour indicates high (low) values for the specific feature. Dots to the right (left) of the vertical line where SHAP value is zero indicate that this feature increases (decreases) predicted probability of cancer diagnosis in year 6. Features written in bold with an asterisk were also among the top 20 in feature importance based on mean absolute SHAP value. (b) XGBoost model average gain value (c) Mean absolute SHAP value for the top 20 features\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-5701032/v1/2ee29569f6fb3bf7a3a04746.png"},{"id":76834188,"identity":"0e294109-2f93-4641-9abe-537a8e4c1e01","added_by":"auto","created_at":"2025-02-21 08:56:51","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":72255,"visible":true,"origin":"","legend":"\u003cp\u003eAn illustration of method A. Cohorts of different sizes are created by applying thresholds to model risk scores. The cancer incidence in such cohorts is calculated and compared to the baseline cancer rate to generate the lift value.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-5701032/v1/94ac33a031cb4fe1e84bf514.png"},{"id":76833141,"identity":"ab1a27a0-aaad-4fc9-855a-9aaf790e95c8","added_by":"auto","created_at":"2025-02-21 08:48:52","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":75499,"visible":true,"origin":"","legend":"\u003cp\u003eVariation in lift value (ratio of cancer incidence in cohort to baseline incidence) with increasing cohort size. Predictions were obtained from a XGBoost model trained on all variables, and on another trained only on demographic variables to showcase the improved accuracy in identifying high risk groups when additional variables such as comorbidities and symptoms related to 111 calls are added.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-5701032/v1/ad9e1e3775f41d7f14624e41.png"},{"id":76834190,"identity":"4c6e496b-d69d-43bc-997b-8af6c8597b24","added_by":"auto","created_at":"2025-02-21 08:56:52","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":129854,"visible":true,"origin":"","legend":"\u003cp\u003eAn illustration of Method B. The most informative features from the trained model are identified and used to identify high-risk cohorts by identifying pairs of features which result in the cohorts with highest incidence. These decision rules can be applied either population wide, or to specific demographic sub-groups.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-5701032/v1/aa4af9651803110248e41b0e.png"},{"id":90344994,"identity":"35404426-55d2-45f8-99c5-ac0d572084bb","added_by":"auto","created_at":"2025-09-01 16:09:02","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1303959,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5701032/v1/aa78773d-c420-4a8c-9e12-e720d239dad8.pdf"},{"id":76833137,"identity":"a88b7a78-3eef-46be-87dc-ce60331361de","added_by":"auto","created_at":"2025-02-21 08:48:51","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":42784,"visible":true,"origin":"","legend":"","description":"","filename":"Constructingmulticancerriskcohortssupplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-5701032/v1/deb5c1184766011f44f72da0.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eConstructing multicancer risk cohorts using national data from medical helplines and secondary care\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eImproving specificity in identifying cohorts at higher risk of developing cancer could increase rates of early diagnosis and allow more focussed interventions to be delivered. However, diagnosing people early is complicated as early-stage symptoms can be harder to definitively attribute to cancer pathology. This means a very large number of individuals would need to be tested to detect a relatively small number of cancers, rendering using such symptoms impractical as a basis for symptomatic case finding or population level screening programmes. To achieve more accurate cancer incidence prediction, the last decade has seen a proliferation of machine learning models trained with unprecedented access to large datasets and computing power. Previous research, in that vein, has typically either i) used routinely collected data, from either secondary or primary care,\u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3 CR4 CR5 CR6 CR7 CR8 CR9\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e or ii) used imaging and/or biomarker/genetic data, which are limited to small segments of the population (e.g. those with specific comorbidities).\u003csup\u003e\u003cspan additionalcitationids=\"CR12 CR13 CR14\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003ePrevious research has provided a wealth of findings highlighting the promise of using machine learning for cancer risk prediction. Both these approaches are not however without their limitations when it comes to early identification of cohorts at higher risk of cancer among the general population. This is because secondary care data may only capture cancer specific events that are picked up quite late in the patient pathway, which could result in worse outcomes. On the other hand, focusing only on primary care data may miss the useful information included in secondary care data, which has shown promise in recent research.\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e Moreover, the use of data that are available for only a small segment of the population is unsuitable for identifying high risk cohorts at the population level.\u003c/p\u003e \u003cp\u003eTo optimise identification of high-risk cohorts, a combination of both primary and secondary care data, at the population level, would be preferable. This approach could capture both symptoms and lifestyle factors, as well as detailed comorbidities, which have previously been shown to be useful signals for predicting future cancer incidence. In the absence of access to national level primary care data, we decided to use National Health Service (NHS) data from medical helplines, specifically, NHS 111 calls data. NHS 111 lines is an alternative source of healthcare advice and information if access to a General Practitioner (GP) is not possible. This dataset records symptoms individuals were concerned about and could provide early insights of undiagnosed cancer, signals which we would miss if relying solely on secondary care data. Our research progresses the field in this direction by matching secondary care data - capturing important features including pre-existing comorbidities, frequency of hospital appointments and demographics - to NHS 111 calls data. This is the first time that data from medical helplines have been used to predict future cancer incidence. Data from NHS 111 calls have the additional advantage that are not \u0026ldquo;clinician-initiated data\u0026rdquo; in the sense that \u0026ldquo;they do not reflect data created through specific actions (or inactions) or insights of the clinician\u0026rdquo;. This is significant because recent research has made a compelling case that predictions based on \u0026ldquo;clinician-initiated data\u0026rdquo; may have limited added value, compared to \u0026ldquo;what the average clinician would decide for the average similar patient\u0026rdquo;.\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eAlongside the rich data from secondary care, we construct detailed patient pathways covering six years of patient history. By doing so, we capture comorbidities and frequency of interactions with the healthcare system. There is some evidence that frequency of interactions in secondary care may reflect the number of missed appointments in primary care, arguably a reflection of patient behaviour, and, as such, including frequency of secondary care interactions could help capture relevant patient behaviour towards their health.\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eWe focus on nine cancers (bladder, head and neck, kidney, lymphoma, myeloma, oesophageal, ovarian, pancreatic, and stomach), which are associated with a high proportion of late-stage diagnoses (stage III and IV) or worsening survival rates in England, and don\u0026rsquo;t currently have screening programmes. We focus on predicting the risk of first cancer diagnosis.\u003c/p\u003e \u003cp\u003eBuilding on model results and making use of feature importance, we successfully develop an approach for constructing higher risk cohorts of varying size while minimising the possible bias that may come from the relatively small numbers of patients for certain demographics. Our approach complements a more standard approach of identifying cohorts at higher risk based on individual risk predictions. It affords greater insights by constructing higher-risk cohorts based on feature importance, the data available to those charged with administering the intervention, and the type of intervention (e.g. whether it targets symptomatic or asymptomatic patients). We illustrate this approach, by applying it to bladder cancer, a priority cancer site.\u003c/p\u003e \u003cp\u003eOur study also provides the first implementation of multi-cancer prediction modelling using population level data in England. Recent work has made the case for the utility of multi-cancer predictive modelling in the context of new liquid biopsy tests, which are currently under development and evaluation.\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e The results presented here could provide further evidence of the possibilities for multi-cancer prediction afforded by national level health data collections.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003ePopulation description\u003c/p\u003e \u003cp\u003eOur dataset includes 23.6\u0026nbsp;million patient histories of individuals between 40 and 74 years old in England (see methods for full details on dataset construction). This age cohort is selected based on the relatively higher incidence of cancer (compared to younger cohorts), and the fact that diagnostics and treatment are less likely to pose complications (e.g. due to frailty), compared to older cohorts. In order to focus on the first cancer diagnosis, we exclude all those with a previous cancer diagnosis from the study population. Our patient histories cover six years between 2016\u0026ndash;2022. We use the first 5 years of the patient histories to predict cancer diagnosis in year 6. A stylised version of a patient pathway diagnosed with cancer in year 6 is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAn individual\u0026rsquo;s patient history includes not only demographic information, but also comorbidities diagnosed during the 5-year period (between 2016 and 2021), as well as information on symptoms reported to NHS 111 lines. Data on mortality allows us to monitor who may have passed away during the prediction period (September 2021-August 2022) for reasons other than cancer and exclude those individuals from our data. We also exclude individuals with a previous cancer diagnosis. More details on dataset construction are included in the methods section.\u003c/p\u003e \u003cp\u003eModel prediction\u003c/p\u003e \u003cp\u003eWe trained several classification models to predict the probability of being diagnosed with cancer in the coming year (September 2021-August 2022). We selected the XGBoost model as our preferred specification based on comparisons in terms of performance with the other classifiers. Given the very sharp class imbalance between cancer and non-cancer cases, we use under sampling in our training datasets to ensure an equal number of cancer and non-cancer cases. We predict the risk of cancer diagnosis for the nine cancer sites selected for the reasons discussed earlier. For each of these models, we report several performance metrics for all cancer specific models (see Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModel performance on the test dataset across different cancer sites.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCancer site\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBladder\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.818\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.731\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.731\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.755\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHead and neck\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.756\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.669\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.669\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.709\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKidney\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.785\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.707\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.707\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.715\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLymphoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.663\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.663\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.705\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMyeloma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.780\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.706\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.706\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.697\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOesophageal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.831\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.707\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.707\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.800\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOvarian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.693\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.625\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.625\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.658\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePancreatic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.809\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.671\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.671\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.796\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStomach\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.792\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.682\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.682\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.752\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eImportant features\u003c/p\u003e \u003cp\u003eIn Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, we present some descriptive statistics focusing on the comparisons between those diagnosed with bladder cancer during the predictive window and those who were not.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDescriptive statistics bladder cancer diagnosis vs no bladder cancer diagnosis in the predictive window (year 6).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBladder cancer diagnosis\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo bladder cancer diagnosis\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCount\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7130\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23620663\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e64.7 (7.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e55.3 (9.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGender (male/female)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e75.9%/24.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e50.8%/49.2%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEthnicity (White British/Other)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e85.9%/14.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e71.6%/28.4%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResiding in 5 most deprived IMD deciles (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e48.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e48.2%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNHS 111 calls reporting cancer related symptoms (during year 5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.17 (0.58)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.05 (0.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eA\u0026amp;E attendances (during year 5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.54 (1.18)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.26 (0.85)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c3\" namest=\"c1\"\u003e \u003cp\u003eIMD stands for Index of Multiple Deprivation. Standard deviation in parenthesis.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAs expected, there are noticeable differences in terms of age and gender between those diagnosed with bladder cancer in year 6 and those who were not. Cancer cases are predominantly male and older, reflecting the well-established link between age and cancer incidence, as well as the fact that bladder cancer is more frequent among males. Beyond demographics, the number of 111 calls reporting cancer related symptoms (see Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e for full list), as well as the number of A\u0026amp;E attendances, are higher on average for those diagnosed with bladder cancer in year 6 compared to those who are not.\u003c/p\u003e \u003cp\u003eBoth the number of calls to NHS 111 lines reporting cancer related symptoms, as well as number of A\u0026amp;E attendances, will end up being among the most useful features for model prediction, as we will see later.\u003c/p\u003e \u003cp\u003eIn order to improve model accuracy and to inform our work on constructing higher risk cohorts, we select the most important features using two metrics, gain and Shapley (SHAP) values. We demonstrate our approach for cohort construction by using bladder cancer as our test case.\u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, we show the SHAP values for the top 20 features (our models include more than 800 features in total) \u0026ndash; ordered based on the gain metric (the average gain across all splits where the feature is used) for the XGBoost model (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). The red colour indicates higher values for the selected feature, and a positive SHAP value means an increase in the risk of cancer. For example, higher age (red colour) has overwhelmingly positive SHAP values, which means that higher age is predicting higher risk of bladder cancer in the next year. By comparison, we show the mean absolute SHAP value for these features (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). While there are slight differences in the ordering of the most informative features, 17 out of the top 20 features based on average model gain are also amongst the top 20 features as determined by mean absolute SHAP value.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe observe that beyond age and gender, several comorbidities appear as relevant predictors of a bladder cancer diagnosis in ways that are consistent with expectations based on the medical literature. For example, the presence of chronic obstructive pulmonary disease (COPD) and urinary infections is associated with the incidence of bladder cancer in previous research.\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eIn addition, several features drawn from the NHS 111 calls dataset appear to be good predictors of bladder cancer incidence. For example, higher number of calls to NHS 111 lines reporting cancer related symptoms is one of the features with the highest gain metric value (just below demographics and long-term condition status). In addition, we also see that features capturing specific symptoms that are plausibly related to undiagnosed bladder cancer are also relevant and have the expected direction of effect. Specifically, higher number of calls to NHS 111 lines reporting \u0026ldquo;pain and frequency of passing urine\u0026rdquo; or \u0026ldquo;blood in urine\u0026rdquo; (during the last year) are both relevant predictors of risk of being diagnosed with bladder cancer in the next year.\u003c/p\u003e \u003cp\u003eTo more comprehensively explore the importance of NHS 111 calls as predictors of risk of future cancer diagnosis, we replicated the analysis based on the gain metric for all other priority cancers beyond bladder. In all cases, features based on NHS 111 calls were among the most influential in predicting future cancer diagnosis. In Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, we report the rank of features, created from NHS 111 calls data, in terms of feature importance based on the gain metric. We do this for all cancer sites included in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eFeature importance rank based on the gain metric for features based on NHS 111 calls.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCancer site\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFeature name (rank)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBladder\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls in last year (4th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls relating to pain/frequency of passing urine in last year (11th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls relating to blood in urine in last year (13th )\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHead and neck\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls in last year (17th )\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKidney\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls related to cancer symptoms in last year (2nd )\u003c/p\u003e \u003cp\u003eNumber of 111 calls in last year (9th )\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLymphoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls in last year (5th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls related to cancer symptoms in last year (10th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls in previous 5 years(13th )\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMyeloma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls in last year (5th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls in previous 5 years (15th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls related to cancer symptoms in last year (17th )\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOesophageal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls in last year (10th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls related to cancer symptoms in last year (11th ),\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOvarian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls in last year (1st )\u003c/p\u003e \u003cp\u003eNumber of 111 calls related to cancer symptoms in last year (5th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls relating to chest and upper back pain in last year (9th )\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePancreatic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls in last year (5th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls related to cancer symptoms in last year (6th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls relating to abdominal pain in last year (12th )\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStomach\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of 111 calls in last year (5th )\u003c/p\u003e \u003cp\u003eNumber of 111 calls related to cancer symptoms in last year (7th )\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e highlights, that features based on information captured in NHS 111 calls are among the top 20 features, based on the gain metric, and often among the top 5 or 10 for all cancer sites we explored in this study.\u003c/p\u003e \u003cp\u003eConstructing higher-risk cohorts\u003c/p\u003e \u003cp\u003eThe primary goal for the analysis is to use the model results to construct high risk cohorts, for a cancer diagnosis within the next year, which can then be used to inform case finding and appropriate interventions to support earlier diagnosis and improve survival. We discuss two possible approaches to achieving this.\u003c/p\u003e \u003cp\u003eIn the first method (Method A), we use the model risk probability at the individual patient level to create cohorts. Different sized cohorts can be constructed by varying the threshold for inclusion in the high-risk group.\u003c/p\u003e \u003cp\u003eThe second method (Method B) identifies cohorts with defined characteristics based on decision rules, utilising the most informative features from the trained model.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eMethod A\u003c/h2\u003e \u003cp\u003eOne approach to constructing higher risk cohorts is to consider capacity based on the requirements of a specific intervention/screening programme, and then selecting the appropriate risk threshold which would lead to the desired cohort size. The risk thresholds are applied to the individual level predictions of the model. Based on different risk thresholds, high-risk cohorts of varying sizes can be constructed. The lower the risk threshold, the larger the size of the cohort, but the lower the potential incidence of cancer within the cohort. An illustration of this method is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eWe define a lift value as the ratio of the cancer incidence within the cohort to the baseline cancer incidence. The baseline cancer incidence in our case, refers to those aged between 40\u0026ndash;74 with no previous cancer diagnoses. Based on different risk thresholds, one could construct a lift curve which plots the lift value against the size of the cohort on the x-axis.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAn example based on the model on bladder cancer is provided in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The lift curve exhibits the expected shape where the lift value declines as we increase the size of the cohort. It also highlights the potential trade-off between high incidence and the total number of cancers correctly predicted. As the cohort size is increased, individuals with lower risk scores are included in the cohort, reducing the incidence and hence the lift value.\u003c/p\u003e \u003cp\u003eTypically, the smaller the selected cohort of the population, the higher the lift value, as these are the individuals with the highest risk scores from the model. For example, in the top left of the figure, considering a cohort size of the highest risk 125,000 individuals (based on model probability risk score), the lift value of the model trained on all variables is 16.6 (representing a cancer rate of 1 in 202 in the cohort). For a model trained on only demographic variables, the lift value for an equivalent cohort size is 9.4 (representing a cancer rate of 1 in 357). These values represent the potential order of magnitude improvement in cancer incidence in identifying high-risk cohorts using a risk score approach compared to chance selection from the population (cancer rate of 1 in 3355). If the cohort size is increased to 1\u0026nbsp;million individuals at highest risk, the lift value reduces to 7.2 for the model with all variables and 5.7 for the model with demographic variables only.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThese results also demonstrate the importance of including features from calls to NHS 111 lines, as well as comorbidities, to the model, as the resulting model risk scores can identify the highest risk individuals more accurately. As the cohort size increases, the difference in lift value between the two reduces, suggesting that both models capture the background demographic risk factors.\u003c/p\u003e \u003cp\u003eWhile more standard in terms of approach, Method A has several limitations from an operational perspective. First, Method A does not allow one to filter on features that may be most useful from a practical perspective. For example, depending on the type of intervention, one may want to focus on symptomatic patients and therefore select cohorts based on specific symptoms. Second, those charged with administering the interventions may not always have access to individual level predictions and instead would have to rely on flags drawing from specific features that are included in the data available to them. For example, eligibility for targeted lung health checks in England relies on age and smoking status.\u003c/p\u003e \u003cp\u003eThis method also relies on applying thresholds to individual patient level predictions. While steps can be taken to explain the model (e.g. through feature importance and other techniques), ultimately the high-risk groups are a heterogeneous cohort. In Method B, we demonstrate how clearly defined cohorts can be created based on specific feature combinations.\u003c/p\u003e \u003cp\u003eA further limitation of this method is that the model may be biased towards predicting higher risk for certain demographic groups, making the high-risk cohort non representative of the actual incidence of bladder cancer in the population. As shown in the SHAP feature importance results (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea), higher age, male, and white ethnicity all tend to increase the model risk score. This does correspond with higher incidence of bladder cancer in this group, however, given the low counts of bladder cancer among other demographic groups, it is difficult to ensure fair representation of all strata when constructing high-risk cohorts using this method, even when steps are taken to balance the training dataset. In the sub-group cohorts of Method B, we show how we sought to address this.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eMethod B\u003c/h3\u003e\n\u003cp\u003eAn alternative method that overcomes the above constraints is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and outlined as follows. First, we select relevant features based on feature importance and data availability. The most important features, which were in the top 20 of model gain and SHAP value, were selected. SHAP was also used to identify the direction of the feature. Features which had a positive impact on the model output (i.e. which tended to increase the risk if the feature was present) were selected.\u003c/p\u003e \u003cp\u003eWe then filter the population based on those features and examine the predicted cancer incidence. As is the case with Method A, we can then compare incidence of cancer in this curated cohort compared to the baseline incidence in the entire population.\u003c/p\u003e \u003cp\u003eThis method has been applied to the whole population, and to sub-groups of demographic strata, demonstrating how the approach can be used for targeted screening.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003ePopulation wide cohorts\u003c/h3\u003e\n\u003cp\u003eThe pair of features which would yield the highest incidence cohorts (on the validation data) of varying size (at least 10,000 to at least 250000) were identified. Subsequently, the selected pair of features was applied to the test data to evaluate the expected bladder cancer incidence in the whole population.\u003c/p\u003e \u003cp\u003eIn Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, we show some examples of such curated cohorts based on combinations of just two features among those that the model considers as high importance for predicting cancer incidence in the next year. The cohort with highest cancer incidence, and a size of at least 10,000, is constructed based on interactions with the 111-call service and includes a specific bladder cancer related symptom of blood in urine. The cancer incidence within this cohort is 41 times higher than the overall incidence in the analysis population, with a cancer incidence of 1 in 82, compared to 1 in 3355 in the study population.\u003c/p\u003e \u003cp\u003eLarger cohorts of high-risk patients are constructed with flags relating to comorbidities of the genitourinary system and other diseases of the urinary system. Applying these flags to the population results in a cohort size of approximately 100,000 individuals, with a cancer rate 6 times higher than in the overall study population.\u003c/p\u003e \u003cp\u003eAn example of a larger cohort of ~\u0026thinsp;290,000 individuals would be constructed by applying the filter of patients having at least one long-term condition, and a diagnosis relating to symptoms and signs involving the genitourinary system in the last 5 years. This results in a cohort with a lift value of 4.5.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExample high-risk cohorts of varying sizes, applied to the whole population. Clearly defined rules, based on the most informative model features, are used to construct the cohorts of varying sizes. Higher cancer incidence is found in smaller cohorts.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature combination\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePopulation size\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncidence in cohort (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLift value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAt least one call reporting cancer related symptoms in last year\u003c/p\u003e \u003cp\u003eAND\u003c/p\u003e \u003cp\u003eat least one call reporting blood in urine in last year\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16,700\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnosis of \u0026ldquo;Symptoms and signs involving the genitourinary system\u0026rdquo; (ICD10 R30-R39) in the last 5 years\u003c/p\u003e \u003cp\u003eAND\u003c/p\u003e \u003cp\u003eDiagnosis of \u0026ldquo;Other diseases of the urinary system\u0026rdquo; (ICD10 N30-N39) in last 5 years\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e98,300\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.18%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHas a long-term condition\u003c/p\u003e \u003cp\u003eAND\u003c/p\u003e \u003cp\u003eDiagnosis of \u0026ldquo;Symptoms and signs involving the genitourinary system\u0026rdquo; (ICD10 R30-R39) in the last 5 years\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e290,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.14%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003eSub-group cohorts\u003c/h3\u003e\n\u003cp\u003eThe population was segmented into demographic groups to investigate if different sets of features can create higher risk cohorts across demographic strata. This was also to address one of the limitations of method A, namely how we can ensure equality of opportunity if model predictions may be biased when there is insufficient training data from all demographic groups.\u003c/p\u003e \u003cp\u003eThe segmentation was based on gender (male/female) and broad ethnicity (White/Non-white), resulting in four groups. Due to the low incidence of bladder cancer, more granular segmentation would have resulted in very small sample sizes.\u003c/p\u003e \u003cp\u003eFor each population segment, the same methodology as described above was applied, with the cohort with the highest incidence of cancer cases being identified. These decision rules were then applied to the test dataset to evaluate the efficacy of the cohort. The lift value was calculated based on the incidence of cancer for each stratum. The results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExample high-risk cohorts applied to demographic strata. For each stratum, the feature combination which resulted in the highest cancer incidence in the cohort (of minimum size 5000) is shown.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDemographic strata\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFeature combination\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCohort size\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eIncidence in cohort (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eLift value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale \u0026ndash; non white ethnicity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAt least 1 A\u0026amp;E attendance in the last year\u003c/p\u003e \u003cp\u003eAND diagnosis of a long term condition\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e195000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.01%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale \u0026ndash; white ethnicity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAt least one call reporting cancer related symptoms in last year\u003c/p\u003e \u003cp\u003eAND at least one call reporting blood in urine in last year\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e47.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale \u0026ndash; non white ethnicity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAt least 1 A\u0026amp;E attendance in the last year\u003c/p\u003e \u003cp\u003eAND diagnosis of COPD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.07%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale \u0026ndash;white ethnicity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAt least one call reporting cancer related symptoms in last year\u003c/p\u003e \u003cp\u003eAND at least one call reporting blood in urine in last year\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7600\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e36\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFor the white ethnicity group, features related to 111 calls are particularly effective in identifying high-risk groups. The specific nature of the symptom information (blood in urine) can result in small cohorts with lift values of 47.5 for white females, and 36 for white males.\u003c/p\u003e \u003cp\u003eIn contrast, for the non-white ethnic group, more general health factors (e.g. A\u0026amp;E attendance) and comorbidities (e.g. COPD) result in the highest risk groups. These cohorts are still significantly higher in cancer incidence compared to baseline rates for these populations, as shown by the lift values of 2.9 for females, and 4.9 for males. However, they are also significantly lower than the lift values obtained for the white ethnic group. This likely reflects health inequalities in the utilisation of services such as 111 calls.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study makes several contributions to the burgeoning literature that seeks to use machine learning to develop useful predictive models for cancer incidence. First, our results demonstrate that information included in medical helplines such as NHS 111 calls contains useful signals predicting a future cancer diagnosis. Our results, show that without exception for the nine cancer types we examined, features based on NHS 111 calls are among the most significant in terms of importance for predicting a future cancer diagnosis. While data quality and coverage are high when it comes to reported symptoms in NHS 111 calls, this dataset is not as comprehensive as primary care datasets. Future work should look to leverage those datasets alongside the information included in secondary care and NHS 111 calls to create a more complete patient history.\u003c/p\u003e \u003cp\u003eThe second contribution of the study is to describe a practical method of constructing higher risk cohorts that could be tailored based on data availability, type of intervention, and desired levels of accuracy. We showcase this approach drawing from the model predictions for bladder cancer incidence. Beyond its greater flexibility, our approach also mitigates the potential for bias due to the underrepresentation of certain demographic groups in the data.\u003c/p\u003e \u003cp\u003eFinally, ours is the first study employing multi-cancer prediction modelling using population level data from England. Our models exhibit good performance for most cancer types. These results further strengthen the case for using routinely collected national health data to risk stratify the population based on risk for future cancer incidence.\u003c/p\u003e \u003cp\u003eThere are numerous potential practical applications of this analysis. Data could be used to inform case finding services for the high-risk cohorts identified. Such case finding services may be used to identify populations at a higher risk of developing cancer, who would benefit from ongoing surveillance, as well as individuals who may warrant an urgent diagnostic test for cancer. Populations with red flag symptoms of cancer, who meet referral thresholds indicated in NG12 (NICE guidelines on suspected cancer), could be triaged directly into urgent suspected cancer pathways.\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e Cohort characteristics could also be used to inform and better target opportunistic cancer checks, as well as local public awareness campaigns, reflecting symptom combinations and/or geographies with increased risk.\u003c/p\u003e \u003cp\u003eOur study is not without its limitations. For one, coverage of early reported symptoms of underlying disease is not as complete as it would have been if we had been able to use data from primary care. Second, predicting future cancer incidence does not differentiate based on stage of cancer diagnosis. It could be argued that predicting cohorts at risk of presenting at a late stage of cancer would be more valuable in terms of improving early diagnosis, as it could allow better targeting of those interventions to such populations. This information is available in the cancer registry database in England but was not available to the authors at the time of this research. Finally, our study does not include any information on lifestyle factors, which almost certainly play an important role in affecting the baseline risk of future cancer incidence. This is a limitation that is common in much of the previous work relying on secondary care data and this study is not going beyond previous research in that regard.\u003c/p\u003e "},{"header":"Methods","content":" \u003cp\u003eDatasets\u003c/p\u003e \u003cp\u003eOur predictive models are trained on a dataset that captures an individual\u0026rsquo;s previous interactions with the healthcare system, their comorbidities, as well as a rich set of socio-demographic information. To create these patient histories, we combine several large datasets including the National Bridges to Health Segmentation Dataset, Secondary Use Services (SUS) data, Emergency Care Data Set (ECDS), NHS 111 calls data, as well as ONS mortality data.\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e A brief description of these datasets is provided below.\u003c/p\u003e \u003cp\u003eThe National Bridges to Health Segmentation Dataset (B2H), which draws on a large number of datasets, provides information on long-term conditions for all patients, over 60\u0026nbsp;million, registered with General Practices in England.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e In addition, we use the information included in B2H for socio-demographic characteristics (e.g. including race, age, sex, deprivation, household type) that could affect the risk of a future cancer diagnosis.\u003c/p\u003e \u003cp\u003eTo complement the information included in B2H, we draw on data from the SUS and ECDS datasets, which include information on all outpatient, inpatient and emergency attendances in hospitals in England. This allows us to capture information on the number of previous inpatients/outpatients and emergency attendances. Frequency of interactions with the health system could reflect attitudes towards one\u0026rsquo;s health, beyond capturing underlying healthcare needs. For this reason, we construct several features which capture separately the number of previous hospital admissions, outpatient appointments, and emergency attendances within different time periods in the past (e.g. last year, last 5 years).\u003c/p\u003e \u003cp\u003eThe SUS/ECDS datasets also allow us to capture detailed comorbidities as diagnosed in secondary care. We use ICD 10 codes covering 263 groups of comorbidities.\u003c/p\u003e \u003cp\u003eA key dataset used in the analysis is the NHS 111 calls dataset. This dataset covers all calls made to NHS 111 lines between 2018\u0026ndash;2023. It includes information on the symptoms the caller reported, as well as the date of the call. With the help of clinicians, we identified symptoms which may be related to cancer and constructed the relevant features capturing both the frequency of any cancer related symptoms reported, as well as the frequency for specific symptoms (e.g. blood in urine). Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e with the cancer related symptoms is below.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCancer related symptoms reported in NHS 111 calls.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"1\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCancer related symptoms reported in NHS 111 calls\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbdominal flank groin or back pain or swelling\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbdominal pain pregnant over 20 weeks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbdominal pain rectal bleeding pregnant over 20 weeks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbdominal pain\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBlood in urine\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBreast lump pregnant\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBreast lump\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBreathing problems breathlessness or wheeze pregnant\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBreathing problems breathlessness or wheeze\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChest and upper back pain\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstipation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCough\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCoughing up blood\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiarrhoea\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDifficulty passing urine\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEasy or unexplained bruising\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eED Triage chest pain\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFace neck pain or swelling\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFever\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGenital problems\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eItch\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMouth ulcers\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePain and/or frequency passing urine\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRectal bleeding\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRectal pain swelling lump or itch\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSkin lumps\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSkin problems\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTiredness fatigue\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUrinary problems\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVaginal bleeding\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVaginal discharge\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVomiting\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVomiting blood\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFinally, to account for censoring due to death, we use person level data from the Office for National Statistics (ONS) death on mortality to capture those passing away during the period.\u003c/p\u003e \u003cp\u003eThe various datasets are linked together using the pseudonymised ID that is common across the datasets. This allows us to create patient histories that capture all patient interactions with secondary care in the NHS, as well as any calls to the NHS 111 lines. Alongside any diagnosed comorbidities and sociodemographic information, this dataset provides rich information on which to build our predictive modelling.\u003c/p\u003e \u003cp\u003eFeature construction and preprocessing\u003c/p\u003e \u003cp\u003eIn the section below, we describe in more detail how the features are constructed. The complete variable list of features used for modelling is listed in Supplementary Table\u0026nbsp;1).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eComorbidity\u003c/h2\u003e \u003cp\u003eFor each ICD10 category 3 code block (e.g. A00-A09, A15-A19 \u0026hellip; Z80-Z99), we create a flag per patient, to indicate if they received a diagnosis in this category, in the last year or last 5 years. This was done by evaluating all diagnosis fields from SUS Inpatient, outpatient, A\u0026amp;E, and the emergency care dataset (ECDS).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eInteractions with the healthcare system\u003c/h3\u003e\n\u003cp\u003eThe number of attendances at A\u0026amp;E, inpatient, and outpatient settings was calculated for each patient in the last year and the last 5 years from the cut-off date. The number of calls to 111 in the last year, and the number of calls with potentially cancer related symptoms was calculated.\u003c/p\u003e\n\u003ch3\u003eSocio-demographic\u003c/h3\u003e\n\u003cp\u003eWe use one-hot encoding for categorical variables (Ethnicity, Index of multiple deprivation, Integrated Care Board, Acorn household type). The latter variables segment households into 6 categories (and 62 types) capturing financial circumstances, benefit receipt, health, wellbeing, and leisure and shopping behaviours.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e In addition we include age, as well as an indicator variable capturing whether the individual is residing in a care home. For modelling purposes, we also impose several exclusions which we discuss below.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eExclusions and missing data\u003c/h2\u003e \u003cp\u003eThe focus of this analysis is individuals who are aged between 40 and 74 at the cut-off date between the observation and prediction window. We therefore exclude younger cohorts who are likely to have a much lower risk of developing cancer, as well as older individuals. We then impose the following restrictions.\u003c/p\u003e \u003cp\u003eWe exclude those with a previous cancer diagnosis, as our focus here is on first diagnosis of cancer. We also exclude those who passed away before the cut-off date as well as those who passed away for any other reason than cancer after the cut-off date. We finally exclude the small number of individuals (1824) with missing information on gender.\u003c/p\u003e \u003cp\u003eWhere data was missing for categorical variables, the null value was replaced with an \u0026lsquo;unknown\u0026rsquo; string value. For those with missing data on ethnicity, we create an \u0026ldquo;unknown\u0026rdquo; flag and include this in the analysis.\u003c/p\u003e \u003cp\u003eMachine learning analysis\u003c/p\u003e \u003cp\u003eThe dataset was split into train (60%), validation (20%), and test (20%) datasets through a random split. We performed a series of statistical tests to examine whether there were still systematic differences between the datasets in terms of demographics. No differences were observed between the datasets (details are included in Supplementary Table\u0026nbsp;3). The size of the datasets is shown in Supplementary Table\u0026nbsp;4.\u003c/p\u003e \u003cp\u003eFor training models, the train dataset was randomly under sampled to ensure an equal number of cancer and non-cancer cases. This was done to avoid the issues stemming from large class imbalance due to the very small incidence of cancer in the data.\u003c/p\u003e \u003cp\u003eWe trained four machine learning models: logistic regression, support vector machine with linear kernel, random forest, and XGBoost. Hyperparameter optimisation was performed by optimising the receiver operating curve area under the curve, using the train and validation datasets with the hyperopt package. The hyperparameters and the ranges for optimisation are provided in the Supplementary Table\u0026nbsp;2. We report model performance using the test dataset. XGBoost appears to perform better compared to the alternatives, a fact consistent with its reputation in terms of performance when it comes to tabular data.\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eAnalysis was performed using python 3.10 on a spark cluster (3.5.0). Versions of the key packages used in the analysis are described in Supplementary Table\u0026nbsp;5.\u003c/p\u003e \u003cp\u003eTraining features\u003c/p\u003e \u003cp\u003eThe XGBoost model was trained with all features (listed in Supplementary Table\u0026nbsp;1) and also with only demographic and socio-economic variables (age, gender, ethnicity, index of multiple deprivation, geographical variable (Integrated Care Board), care home flag, and acorn household type in order to explore the impact of features relating to 111 calls, comorbidities, and healthcare interactions on the model performance and high-risk cohorts.\u003c/p\u003e \u003cp\u003eFeature importance\u003c/p\u003e \u003cp\u003eModel feature importance was obtained from the XGBoost model by ranking features by their average gain across all splits the feature is used in. SHAP values were calculated on the validation dataset from trained models. SHAP calculates the contribution of each variable to the model predicted probability output.\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eCohort construction\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eMethod A\u003c/h2\u003e \u003cp\u003eIndividual patient level predictions were obtained on the test dataset. High-risk cohorts were constructed by varying the risk threshold and evaluating the cancer incidence within the cohort.\u003c/p\u003e \u003cp\u003eThe cohort size (as shown in the lift curve in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) was obtained by extrapolating to the whole population from the test dataset (which is a random sample comprising 20% of the whole study population).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eMethod B\u003c/h2\u003e \u003cp\u003eThe top 20 most informative features from model gain and SHAP were identified. Features which were present in both lists, and which tended to increase the risk if the feature was present, were selected. Demographic (gender and ethnicity) features were not selected as they were used to segment the population in the sub-group cohorts.\u003c/p\u003e \u003cp\u003eEach pair of selected features was used to filter the validation dataset. The size and incidence in the resulting cohort were calculated.\u003c/p\u003e \u003cp\u003eFor a particular cohort size, the combination of features which resulted in the highest cancer incidence was identified. This pairing of features was then applied to the unseen hold out test dataset to calculate the expected cancer incidence in the wider population. The cohort size in the whole study population was obtained by extrapolating from the test dataset.\u003c/p\u003e \u003cp\u003eFor sub-groups, the same process as above was applied, with the difference that an additional filtering of the data by demographic strata was also applied.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eEthical Approval\u003c/h2\u003e \u003cp\u003eNot applicable. Data is collected and used in line with NHS England\u0026rsquo;s purposes as required under the statutory duties outlined in the NHS Act 2006 and Health and Social Care Act 2012. Data is processed using best practice methodology underpinned by a Data Processing Agreement between NHS England and Outcomes Based Healthcare Ltd (OBH), who produce the Segmentation Dataset on behalf of NHS England. This ensures controlled access by appropriate approved individuals, to anonymised/pseudonymised data held on secure data environments entirely within the NHS England infrastructure. Data is processed for specific purposes only, including operational functions, service evaluation, and service improvement. Where OBH has processed data, this has been agreed and is detailed in a Data Processing Agreement. The data used to produce this analysis has been disseminated to NHS England under Directions issued under Section 254 of the Health and Social Care Act 2012.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eD.P. and H.M. contributed equally as co-first authors to this work. D.P. wrote the main manuscript text and contributed to the design of the analytical methodology. H.M. led on the design of the analytical methodology, contributed to the writing of the main text/supplementary material and to the analysis. D.B. led on the design of the analytical methodology, contributed to the writing of the main text and to the analysis. S.K. contributed to the analysis and to the writing of the main text/supplementary material. G.T., R.Ch., A.M. contributed to the scoping of the analysis, the design of the analytical methodology and manuscript revisions. R.Ca., E.H.W. contributed to the scoping of the analysis and the writing of the main manuscript. All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe authors would like to thank Robert Scott, and Thomas Henstock for their contributions in the scoping and analysis at the early stages of this study. We would also like to thank Michael Spence and Rajun Phagura for their help at various stages of the analysis. Finally, we are grateful to Anthony Cunliffe GP, Afsana Bhuyia GP, Tina George GP and Amelia Randle GP for sharing their clinical expertise.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eAll data used in this study are held internally by NHS England. The data cannot be shared publicly as they contain patient level sensitive information.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eReporting summary\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFurther information on research design is available in the Nature Research Reporting Summary linked to this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCODE AVAILABILITY\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLink to data processing notebook: https://github.com/nhsengland/cancer_foundry_data_modelling/\u003c/p\u003e\n\u003cp\u003eCode for data modelling available upon request from the authors. We plan to publish this code in the near future.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAppelbaum, L., et al., 2021. Development and validation of a pancreatic cancer risk model for the general population using electronic health records: An observational study. European Journal of Cancer, 143, pp.19\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, Y.H., Nguyen, P.A., Islam, M.M., Li, Y.C. and Yang, H.C., 2019. Development of deep learning algorithm for detection of colorectal cancer in EHR data. In: \u003cem\u003eMEDINFO 2019: Health and Wellbeing e-Networks for All\u003c/em\u003e. IOS Press, pp.438\u0026ndash;441.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, X., et al., 2019. Prediction of the 1-year risk of incident lung cancer: Prospective study using electronic health records from the State of Maine. Journal of Medical Internet Research, 21(5), p.e13260.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHippisley-Cox, J. and Coupland, C., 2015. Development and validation of risk prediction algorithms to estimate future risk of common cancers in men and women: Prospective cohort study. BMJ Open, 5, p.e007825.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMalhotra, A., et al., 2021. Can we screen for pancreatic cancer? Identifying a sub-population of patients at high risk of subsequent diagnosis using machine learning techniques applied to primary care data. PLOS ONE, 16(6), p.e0251876.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBriggs, E., et al., 2022. Machine learning for risk prediction of oesophago-gastric cancer in primary care: Comparison with existing risk-assessment tools. \u003cem\u003eCancers\u003c/em\u003e, 14, p.5023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIoannou, G.N., et al. 2018. Development of models estimating the risk of hepatocellular carcinoma after antiviral treatment for hepatitis C. Journal of Hepatology, 69(5), pp.1088\u0026ndash;1098.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIoannou, G.N., et al., 2020. Assessment of a deep learning model to predict hepatocellular carcinoma in patients with hepatitis C cirrhosis. JAMA Network Open, 3(9), p.e2015626.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHowell, D., et al., 2023. Developing a risk prediction tool for lung cancer in Kent and Medway, England: Cohort study using linked data. \u003cem\u003eBJC Reports\u003c/em\u003e, 1, p.16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhen, J., et al., 2024. Development and validation of machine learning models for young-onset colorectal cancer risk stratification. \u003cem\u003enpj Precision Oncology\u003c/em\u003e, 8, p.239.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSteinbuss, G., et al. 2021. Deep learning for the classification of non-Hodgkin lymphoma on histopathological images. \u003cem\u003eCancers\u003c/em\u003e, 13(10), p.2419.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaur, L., Bhandari, M., Razdan, T., Mallik, S. and Zhao, Z., 2022. Explanation-driven deep learning model for prediction of brain tumour status using MRI image data. \u003cem\u003eFrontiers in Genetics\u003c/em\u003e, 13, p.822666.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFontanillas, P., et al., 2021. Disease risk scores for skin cancers. \u003cem\u003eNature Communications\u003c/em\u003e, 12, p.160.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTammem\u0026auml;gi, M.C., et al., 2019. Development and validation of a multivariable lung cancer risk prediction model that includes low-dose computed tomography screening results: A secondary analysis of data from the National Lung Screening Trial. JAMA Network Open, 2, p.e190204.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVarma, A. et al., 2023. Early prediction of prostate cancer risk in younger men using polygenic risk scores and electronic health records. Cancer Med, 12, pp. 379\u0026ndash;386.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlacido, D., et al., 2023. A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories. Nature Medicine, 29, pp.1113\u0026ndash;1122.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeaulieu-Jones, B.K., et al., 2021. Machine learning for patient risk stratification: Standing on, or looking over, the shoulders of clinicians? npj Digital Medicine, 4, p.62.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilliamson, A.E., McQueenie, R., Ellis, D.A., McConnachie, A. and Wilson, P., 2021. \u0026lsquo;Missingness\u0026rsquo; in health care: Associations between hospital utilization and missed appointments in general practice. A retrospective cohort study. PLOS ONE, 16(6), p.e0253163.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJung, A.W., Holm, et al. 2024. Multi-cancer risk stratification based on national health data: A retrospective modelling and validation study. The Lancet Digital Health, 6(6), pp.e396-e406.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShephard, E.A., et al. 2012. Clinical features of bladder cancer in primary care. British Journal of General Practice, 62(602), pp.e598-e604.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiao, S., Wang, Y., Zhou, J., Liu, Y., He, S., Zhang, L., Liu, M., Wen, D., Sun, P., Lu, G., Wang, Q., Ouyang, Y. and Song, Y., 2024. Associations between chronic obstructive pulmonary disease and ten common cancers: Novel insights from Mendelian randomization analyses. \u003cem\u003eBMC Cancer\u003c/em\u003e, 24(1), p.601.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNational Institute for Health and Care Excellence. \u003cem\u003eNice guideline, NG12: Suspected cancer: recognition and referral\u003c/em\u003e, last updated 2 October 2023, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.nice.org.uk/guidance/ng12\u003c/span\u003e\u003cspan address=\"https://www.nice.org.uk/guidance/ng12\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNHS Digital. \u003cem\u003eSecondary Uses Services (SUS)\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://digital.nhs.uk/services/secondary-uses-service-sus\u003c/span\u003e\u003cspan address=\"https://digital.nhs.uk/services/secondary-uses-service-sus\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNHS Digital. \u003cem\u003eEmergency Care Data Set (ECDS)\u003c/em\u003e. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://digital.nhs.uk/data-and-information/data-collections-and-data-sets/data-sets/emergency-care-data-set-ecds\u003c/span\u003e\u003cspan address=\"https://digital.nhs.uk/data-and-information/data-collections-and-data-sets/data-sets/emergency-care-data-set-ecds\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eValabhji, J., et al., 2024. Prevalence of multiple long-term conditions (multimorbidity) in England: A whole population study of over 60 million people. Journal of the Royal Society of Medicine, 117(3), pp.104\u0026ndash;117.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAcorn, 2013. The Household Acorn User Guide. Available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.caci.co.uk/wp-content/uploads/2021/07/Household_Acorn_UG.pdf\u003c/span\u003e\u003cspan address=\"https://www.caci.co.uk/wp-content/uploads/2021/07/Household_Acorn_UG.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e [Accessed 16 December 2024].\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBorisov, V., et al. 2022. Deep neural networks and tabular data: A survey. IEEE Transactions on Neural Networks and Learning Systems.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg, S.M. and Lee, S.I., 2017. A unified approach to interpreting model predictions. In: \u003cem\u003eProceedings of the 31st International Conference on Neural Information Processing Systems\u003c/em\u003e. Curran Associates Inc., pp.4768\u0026ndash;4777.\u003c/span\u003e\u003c/li\u003e\u003c/ol\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":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-5701032/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5701032/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIdentification of cohorts at higher risk of cancer can enable earlier diagnosis of the disease, which significantly improves patient outcomes. In this study, we use machine learning to predict cancer diagnosis in the next year. We select nine cancer sites with high incidence of late-stage diagnosis or worsening survival rates, and where there are currently no national screening programmes. We use National Health Service (NHS) data from medical helplines (NHS 111) and secondary care appointments from all hospitals in England. We show that features based on information captured in NHS 111 calls are among the most influential in driving predictions of a future cancer diagnosis. Our predictive models exhibit good discrimination (AUC \u0026ndash; 0.78 \u0026ndash; SD 0.04), ranging from 0.69 (ovarian cancer) to 0.83 (oesophageal cancer). While our predictive modelling provides patient level risk predictions, our emphasis is on constructing cohorts of patients who may be at risk of cancer rather than individual risk scores. We present an approach of constructing cohorts at higher risk of cancer based on feature importance and considering possible bias in model results. These outputs can be used to develop highly targeted case finding services, which could help increase earlier detection rates and reduce health disparities. This approach is flexible and can be tailored based on the group the intervention targets (i.e. symptomatic or asymptomatic patients) and the data available to those charged with administering the intervention.\u003c/p\u003e","manuscriptTitle":"Constructing multicancer risk cohorts using national data from medical helplines and secondary care","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-02-21 08:48:47","doi":"10.21203/rs.3.rs-5701032/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-04-09T03:56:04+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-03-22T23:04:33+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"272580598898199567963487554700096796829","date":"2025-03-09T18:20:44+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"246653160532036638509958040146370862905","date":"2025-03-07T20:04:29+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-02-26T23:33:48+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"57214921196144080093729638849029171489","date":"2025-02-11T17:40:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"239862636443836161691172433410331780733","date":"2025-01-21T08:48:15+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-01-18T16:31:02+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-12-31T01:05:12+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-12-30T05:30:13+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Digital Medicine","date":"2024-12-23T16:02:30+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"88af1d66-d810-4757-bc5f-e3ec5823e8cd","owner":[],"postedDate":"February 21st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":44575524,"name":"Physical sciences/Mathematics and computing/Statistics"},{"id":44575525,"name":"Biological sciences/Cancer"},{"id":44575526,"name":"Health sciences/Health care"}],"tags":[],"updatedAt":"2025-09-01T16:03:38+00:00","versionOfRecord":{"articleIdentity":"rs-5701032","link":"https://doi.org/10.1038/s41746-025-01855-0","journal":{"identity":"npj-digital-medicine","isVorOnly":false,"title":"npj Digital Medicine"},"publishedOn":"2025-08-27 15:58:11","publishedOnDateReadable":"August 27th, 2025"},"versionCreatedAt":"2025-02-21 08:48:47","video":"","vorDoi":"10.1038/s41746-025-01855-0","vorDoiUrl":"https://doi.org/10.1038/s41746-025-01855-0","workflowStages":[]},"version":"v1","identity":"rs-5701032","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5701032","identity":"rs-5701032","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}