Background

The clinical use of personalised prediction models in psychiatry is becoming increasingly feasible. Much research has focused on their use in early intervention in psychosis (in particular, to predict the likelihood of transition to psychosis in individuals at increased risk or to predict outcomes in first-episode psychosis), but it is envisaged that they could be applied to a wide range of psychiatric illnesses occurring across the lifespan.^{Reference Leighton, Upthegrove, Krishnadas, Benros, Broome and Gkoutos1} Published examples include models predicting the individual likelihood of diagnosis of autism spectrum disorder; of the onset of bipolar disorders in individuals at familial risk; and of treatment response to antidepressants in major depression.^{Reference de Pablo, Studerus, Vaquerizo-Serrano, Irving and Catalan2}

Practical considerations

Why do personalised prediction models warrant ethical scrutiny? After all, we routinely make informed estimates of patients’ likely outcomes in clinical practice, based on the available evidence and professional experience. However, personalised prediction models differ from these conventional means in two salient ways: first, their reported predictive accuracy may overstate that achievable in clinical practice; and second, their workings often lack transparency.

In brief, it has been widely observed that personalised prediction models tend to perform more poorly when applied to a different population from the one in which they are derived. However, most personalised prediction models under development for use in psychiatry have not yet undergone external validation using an independent sample or had their accuracy prospectively assessed in a real-world patient population. Therefore, the reported predictive accuracy of currently available models is likely to overestimate what they can achieve in clinical psychiatric practice.^{Reference de Pablo, Studerus, Vaquerizo-Serrano, Irving and Catalan2}

Moreover, current personalised prediction models often lack transparency in terms of not supplying the user with an explanation of how or why a particular risk estimate is made. In addition, the complexity of statistical and machine learning methods makes it extremely difficult (and often impossible) for the clinician, patient or relative to independently decipher the prediction model's workings.^{Reference Grote and Berens3}

Ethical implications

Issues of accuracy and transparency are of ethical importance as they determine the degree of influence that personalised predictions are likely to exert on clinical decision-making. In effect, if prediction models claim a heightened level of predictive accuracy, they are likely to be ascribed greater weight in determining clinical recommendations. Moreover, if clinicians cannot access how or why a particular risk estimate has been made, they are unable to properly evaluate the relevance of it in guiding clinical management. This inability to effectively question risk estimates, combined with their high level of reported accuracy, may leave clinicians feeling pressurised into adopting an automaticity of decision-making, whereby they act in accordance with prediction model outcomes by default.^{Reference Grote and Berens3}

Within psychiatry, this effect could be particularly problematic in high-risk scenarios (e.g. crisis management and in-patient discharge). Typically, such situations are approached through shared decision-making and safety planning with the patient and multidisciplinary colleagues, with preference given to mitigating risk in the least restrictive manner. However, if a prediction model confidently estimates a high chance of a poor outcome, this could be taken to outweigh all other considerations, not least owing to the challenge of justifying acting contrary to this information and fear of medico-legal repercussions if one were to do so. Moreover, the lack of transparency of prediction models renders clinicians ill-equipped to discuss in detail with patients the rationale behind clinical recommendations, thus jeopardising their ability to provide valid informed consent. Therefore, instead of being a useful decision-making adjunct, prediction models could, in some instances, subvert the roles of the healthcare team and patient in shared decision-making, encourage defensive medicine at the expense of patient autonomy and undermine the principle of informed consent.^{Reference Grote and Berens3}

Prediction models could further disempower patients by undermining their sense of agency (the belief that one can shape one's own life).^{Reference Houlders, Bortolotti and Broome4} This may occur if patients misconstrue personalised predictions to infer that significant outcomes (e.g. onset of mental illness, relapse or recovery) are entirely foreseeable and thus entirely predetermined and beyond their ability to influence. This could leave patients feeling hopeless in the face of negative predictions or disengaged in response to positive ones. In a clinical context, patients may be less motivated to play an active role in shared decision-making and treatment, to the detriment of clinical outcomes.^{Reference Houlders, Bortolotti and Broome4} More broadly, patients (and relatives) may impose limitations on their own aspirations for the future, thus curtailing their life experiences. Further, discrimination from external sources could occur if third parties gain access to risk estimates, for example reduced opportunities for employment in certain sectors (e.g. the military) and difficulty obtaining insurance or travel visas.^{Reference Lane, Hunter and Lawrie5}

Last, the clinical use of personalised prediction models risks perpetuating inequities in psychiatric care. Specifically, the real-world accuracy of predictions is likely to be further compromised in individuals from minority groups, owing to underrepresentation in the derivation sample on which the prediction model is trained and tested. For example, if a model is derived from a predominantly young adult, White male population (as is typical of research cohorts in the UK), its risk predictions may be less accurate for individuals from ethnic minorities, females or those at the extremes of age, thus disadvantaging these groups.^{Reference Grote and Berens3}

Ethical recommendations

It is imperative that the above ethical considerations are taken into account in efforts to implement personalised prediction models in clinical psychiatry. In support of this, we suggest the following practical ethical recommendations and highlight avenues for future research.

First, a realistic report of the predictive accuracy achievable by a prediction model in clinical practice should be made available to users, to allow them to make an informed judgement on the influence that risk estimates should have on decision-making. This necessitates that models undergo external validation at the research stage, and prospective assessment of real-world accuracy in a clinical context as part of implementation research, prior to widespread roll-out. Care must be taken to ensure that sample populations at each stage are diverse and inclusive, to avoid creating inequities in predictive accuracy that perpetuate disadvantage for underrepresented groups.

Second, the interpretability of prediction models should be maximised before clinical implementation. Developers should prioritise simpler models where doing so will not compromise clinical utility, for example those using basic clinical data rather than complex biomarkers, and standard statistical approaches rather than sophisticated machine learning algorithms. Initial indications from psychiatric research suggest that simpler models display comparable predictive accuracy to more complex alternatives.^{Reference de Pablo, Studerus, Vaquerizo-Serrano, Irving and Catalan2} In addition, materials should be designed to educate clinicians on how the prediction model operates. In line with this, future research should investigate what level of detail healthcare professionals deem useful. This is likely to include, at a minimum, information on the underlying data-sets and relevant variables used, broadly how the results are derived, how outcomes are defined and, crucially, the reliability and limitations of predictions.

Third, patients must be supported to meaningfully understand their personalised predictions. This is contingent on the previous point of clinicians being adequately informed to educate their patients (and by doing so, facilitate informed consent for interventions based on risk estimates). However, it necessitates not just that clinicians know what information to impart, but also how to do so in a manner that instils hope and maximises agency.^{Reference Lane, Hunter and Lawrie5} It should be made explicit that predictions are probabilistic estimates, as opposed to binary indicators that an outcome will or will not occur. In addition, it should be explained that an outcome as defined by the model (e.g. occurrence of a condition based on diagnostic criteria) does not equate to a uniform lived experience. Last, it must be emphasised that the individual can often influence outcomes, for example by addressing modifiable risk factors and engaging with treatment and support. Future empirical research involving patients is warranted to investigate how best to disclose predictive information in an empowering fashion.