The literature on artificial intelligence (AI) or machine learning (ML) in mental health and psychiatry lacks consensus on what “explainability” means. In the more general XAI (eXplainable AI) literature, there has been some convergence on explainability meaning model-agnostic techniques that augment a complex model (with internal mechanics intractable for human understanding) with a simpler model argued to deliver results that humans can comprehend. Given the differing usage and intended meaning of the term “explainability” in AI and ML, we propose instead to approximate model/algorithm explainability by understandability defined as a function of transparency and interpretability. These concepts are easier to articulate, to “ground” in our understanding of how algorithms and models operate and are used more consistently in the literature. We describe the TIFU (Transparency and Interpretability For Understandability) framework and examine how this applies to the landscape of AI/ML in mental health research. We argue that the need for understandablity is heightened in psychiatry because data describing the syndromes, outcomes, disorders and signs/symptoms possess probabilistic relationships to each other—as do the tentative aetiologies and multifactorial social- and psychological-determinants of disorders. If we develop and deploy AI/ML models, ensuring human understandability of the inputs, processes and outputs of these models is essential to develop trustworthy systems fit for deployment.

Artificial intelligence (AI) can help clinicians to improve healthcare decision-making by integrating data from high-volume, heterogeneous electronic health records (EHRs). However, there is growing evidence that AI solutions in healthcare carry considerable risk of harm for people belonging to racial, ethnic, sexual and gender minority communities, which can exacerbate inequalities.

Accessing specialist secondary mental health care in the NHS in England requires a referral, usually from primary or acute care. Community mental health teams triage these referrals deciding on the most appropriate team to meet patients’ needs. Referrals require resource-intensive review by clinicians and often, collation and review of the patient’s history with services captured in their electronic health records (EHR). Triage processes are, however, opaque and often result in patients not receiving appropriate and timely access to care that is a particular concern for some minority and under-represented groups. Our project, funded by the National Institute of Health Research (NIHR) will develop a clinical decision support tool (CDST) to deliver accurate, explainable and justified triage recommendations to assist clinicians and expedite access to secondary mental health care

Validated instruments such as questionnaires, patient-reported outcome measures and clinician-rated psychopathology scales, are indispensable for measuring symptom burden and mental state, and for defining outcomes in both psychiatric practice and clinical trials. Most often, the values on the instrument’s multiple items (dimensions) are added to derive a single, univariate (scalar) sum-score. Although this approach simplifies interpretation, there are always many possible combinations of individual items that can yield the same sum-score. Two patients can therefore obtain identical scores on a given instrument, despite having very different combinations of underlying item scores corresponding to different patterns of clinical symptoms. The same is also true when a single patient is measured at two different time points, where the resulting sum-scores can obscure changes that may be clinically meaningful. We present an alternative analytic framework, which leverages geometric concepts to represent measurements as points in a vector space. Using this framework, we show why sum-scores obscure information present in measurements of clinical state, and also provide a straightforward algorithm to mitigate against this problem. Clinically-relevant outcomes, such as remission or patient-centered treatment goals, can be represented intuitively, as reference points or ‘anchors’ within this space. Using real-world data, we then demonstrate how measuring the relative distance between points and anchors preserves more information, allowing outcomes such as proximity to remission, to be defined and measured.

Sleep disruption is a common precursor to deterioration and relapse in people living with psychotic disorders. Understanding the temporal relationship between sleep and psychopathology is important for identifying and developing interventions which target key variables that contribute to relapse.

The last decade’s growth in artificial intelligence, machine learning, and statistical methods for high-dimensional data has driven a zeitgeist of prediction (or forecasting) in medicine and psychiatry. Algorithms for prediction require a model that is governed by parameters whose values are estimated from exemplar training cases. Estimation (or training) of parameters ingrains uncertainty into the resulting algorithm arising from model assumptions in addition to bias and error in the data. The trained algorithm’s proficiency is tested on separate validation cases (not seen during training) and summarized as representative of the expected performance when used for making predictions about actual patients. The trained model yields a continuous score that is proportional to the probability of some outcome, commonly a diagnosis or the occurrence of an event. Most often, this continuous score is compared with an operating threshold (or cutoff) that implicitly defines a dichotomizing decision rule because this is compatible with summary measures of performance (SMP) such as the area under the receiver operating characteristic curve (AUROC), sensitivity/specificity, and balanced accuracy. Sometimes, the continuous scores are instead summarized as the Brier score, ranging from 0 (perfect) to 1 (worst). In this Viewpoint, we discuss an important but neglected issue: summary measures of performance obscure uncertainty in the algorithm’s predictions that may be relevant when deployed for clinical decision-making.

BACKGROUND: Stratified or personalised medicine targets treatments for groups of individuals with a disorder based on individual heterogeneity and shared factors that influence the likelihood of response. Psychiatry has traditionally defined diagnoses by constellations of co-occurring signs and symptoms that are assigned a categorical label (e.g. schizophrenia). Trial methodology in psychiatry has evaluated interventions targeted at these categorical entities, with diagnoses being equated to disorders. Recent insights into both the nosology and neurobiology of psychiatric disorder reveal that traditional categorical diagnoses cannot be equated with disorders. We argue that current quantitative methodology (1) inherits these categorical assumptions, (2) allows only for the discovery of average treatment response, (3) relies on composite outcome measures and (4) sacrifices valuable predictive information for stratified and personalised treatment in psychiatry. METHODS AND FINDINGS: To achieve a truly ‘stratified psychiatry’ we propose and then operationalise two necessary steps: first, a formal multi-dimensional representation of disorder definition and clinical state, and second, the similar redefinition of outcomes as multidimensional constructs that can expose within- and between-patient differences in response. We use the categorical diagnosis of schizophrenia-conceptualised as a label for heterogeneous disorders-as a means of introducing operational definitions of stratified psychiatry using principles from multivariate analysis. We demonstrate this framework by application to the Clinical Antipsychotic Trials of Intervention Effectiveness dataset, showing heterogeneity in both patient clinical states and their trajectories after treatment that are lost in the traditional categorical approach with composite outcomes. We then systematically review a decade of registered clinical trials for cognitive deficits in schizophrenia highlighting existing assumptions of categorical diagnoses and aggregate outcomes while identifying a small number of trials that could be reanalysed using our proposal. CONCLUSION: We describe quantitative methods for the development of a multi-dimensional model of clinical state, disorders and trajectories which practically realises stratified psychiatry. We highlight the potential for recovering existing trial data, the implications for stratified psychiatry in trial design and clinical treatment and finally, describe different kinds of probabilistic reasoning tools necessary to implement stratification.