Explainable artificial intelligence models for predicting risk of suicide using health administrative data in Quebec.

Suicide is a complex, multidimensional event, and a significant challenge for prevention globally. Artificial intelligence (AI) and machine learning (ML) have emerged to harness large-scale datasets to enhance risk detection. In order to trust and act upon the predictions made with ML, more intuitive user interfaces must be validated. Thus, Interpretable AI is one of the crucial directions which could allow policy and decision makers to make reasonable and data-driven decisions that can ultimately lead to better mental health services planning and suicide prevention. This research aimed to develop sex-specific ML models for predicting the population risk of suicide and to interpret the models. Data were from the Quebec Integrated Chronic Disease Surveillance System (QICDSS), covering up to 98% of the population in the province of Quebec and containing data for over 20,000 suicides between 2002 and 2019. We employed a case-control study design. Individuals were considered cases if they were aged 15+ and had died from suicide between January 1st, 2002, and December 31st, 2019 (n = 18339). Controls were a random sample of 1% of the Quebec population aged 15+ of each year, who were alive on December 31st of each year, from 2002 to 2019 (n = 1,307,370). We included 103 features, including individual, programmatic, systemic, and community factors, measured up to five years prior to the suicide events. We trained and then validated the sex-specific predictive risk model using supervised ML algorithms, including Logistic Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Multilayer perceptron (MLP). We computed operating characteristics, including sensitivity, specificity, and Positive Predictive Value (PPV). We then generated receiver operating characteristic (ROC) curves to predict suicides and calibration measures. For interpretability, Shapley Additive Explanations (SHAP) was used with the global explanation to determine how much the input features contribute to the models' output and the largest absolute coefficients. The best sensitivity was 0.38 with logistic regression for males and 0.47 with MLP for females; the XGBoost Classifier with 0.25 for males and 0.19 for females had the best precision (PPV). This study demonstrated the useful potential of explainable AI models as tools for decision-making and population-level suicide prevention actions. The ML models included individual, programmatic, systemic, and community levels variables available routinely to decision makers and planners in a public managed care system. Caution shall be exercised in the interpretation of variables associated in a predictive model since they are not causal, and other designs are required to establish the value of individual treatments. The next steps are to produce an intuitive user interface for decision makers, planners and other stakeholders like clinicians or representatives of families and people with live experience of suicidal behaviors or death by suicide. For example, how variations in the quality of local area primary care programs for depression or substance use disorders or increased in regional mental health and addiction budgets would lower suicide rates.

A case-control study on predicting population risk of suicide using health administrative data: a research protocol.

INTRODUCTION: Suicide has a complex aetiology and is a result of the interaction among the risk and protective factors at the individual, healthcare system and population levels. Therefore, policy and decision makers and mental health service planners can play an important role in suicide prevention. Although a number of suicide risk predictive tools have been developed, these tools were designed to be used by clinicians for assessing individual risk of suicide. There have been no risk predictive models to be used by policy and decision makers for predicting population risk of suicide at the national, provincial and regional levels. This paper aimed to describe the rationale and methodology for developing risk predictive models for population risk of suicide. METHODS AND ANALYSIS: A case-control study design will be used to develop sex-specific risk predictive models for population risk of suicide, using statistical regression and machine learning techniques. Routinely collected health administrative data in Quebec, Canada, and community-level social deprivation and marginalisation data will be used. The developed models will be transformed into the models that can be readily used by policy and decision makers. Two rounds of qualitative interviews with end-users and other stakeholders were proposed to understand their views about the developed models and potential systematic, social and ethical issues for implementation; the first round of qualitative interviews has been completed. We included 9440 suicide cases (7234 males and 2206 females) and 661 780 controls for model development. Three hundred and forty-seven variables at individual, healthcare system and community levels have been identified and will be included in least absolute shrinkage and selection operator regression for feature selection. ETHICS AND DISSEMINATION: This study is approved by the Health Research Ethnics Committee of Dalhousie University, Canada. This study takes an integrated knowledge translation approach, involving knowledge users from the beginning of the process.

The 5-year longitudinal diagnostic profile and health services utilization of patients treated with electroconvulsive therapy in Quebec: a population-based study.

PURPOSE: Electroconvulsive therapy (ECT) is effective for treating several psychiatric disorders. However, only a minority of patients are treated with ECT. It is of primary importance to characterize their profile for epidemiological purposes and to inform clinical practice. We aimed to characterize the longitudinal profile of psychopathology and services utilization of patients first treated with ECT. METHODS: We conducted a population-based comparative study using data from a national administrative database in Quebec. Patients who received a first ECT between 2002 and 2016 were compared to controls who were hospitalized in psychiatry but did not receive ECT. We performed descriptive analyses to compare psychiatric diagnoses, domains of psychopathology (internalizing, externalizing and thought/psychotic disorders), medical services and medication use in the 5 years prior to the ECT or hospitalization. RESULTS: 5 080 ECT patients were compared with 179 594 controls. Depressive, anxiety, bipolar and psychotic disorders were more frequent in the ECT group. 96.2% of ECT patients had been diagnosed with depression and 53.8% with a primary psychotic disorder. In the ECT group, 1.0% had been diagnosed exclusively with depression and 47.0% had disorders from that belong to all three domains of psychopathology. Having both internalizing and thought/psychotic disorders was associated with an increased likelihood of receiving ECT vs having internalizing disorders alone (unadjusted OR = 2.93; 95% CI = 2.63, 3.26). All indicators of mental health services utilization showed higher use among ECT patients. CONCLUSION: Our results provide robust evidence of complex longitudinal psychopathology and extensive services utilization among ECT patients.

Feature sensitivity criterion-based sampling strategy from the Optimization based on Phylogram Analysis (Fs-OPA) and Cox regression applied to mental disorder datasets.

Digital datasets in several health care facilities, as hospitals and prehospital services, accumulated data from thousands of patients for more than a decade. In general, there is no local team with enough experts with the required different skills capable of analyzing them in entirety. The integration of those abilities usually demands a relatively long-period and is cost. Considering that scenario, this paper proposes a new Feature Sensitivity technique that can automatically deal with a large dataset. It uses a criterion-based sampling strategy from the Optimization based on Phylogram Analysis. Called FS-opa, the new approach seems proper for dealing with any types of raw data from health centers and manipulate their entire datasets. Besides, FS-opa can find the principal features for the construction of inference models without depending on expert knowledge of the problem domain. The selected features can be combined with usual statistical or machine learning methods to perform predictions. The new method can mine entire datasets from scratch. FS-opa was evaluated using a relatively large dataset from electronic health records of mental disorder prehospital services in Brazil. Cox's approach was integrated to FS-opa to generate survival analysis models related to the length of stay (LOS) in hospitals, assuming that it is a relevant aspect that can benefit estimates of the efficiency of hospitals and the quality of patient treatments. Since FS-opa can work with raw datasets, no knowledge from the problem domain was used to obtain the preliminary prediction models found. Results show that FS-opa succeeded in performing a feature sensitivity analysis using only the raw data available. In this way, FS-opa can find the principal features without bias of an inference model, since the proposed method does not use it. Moreover, the experiments show that FS-opa can provide models with a useful trade-off according to their representativeness and parsimony. It can benefit further analyses by experts since they can focus on aspects that benefit problem modeling.