Risk stratification versus precision psychiatry/medicine

The commentary’s key rhetorical move is to treat risk stratification as precision medicine. The implicit claim is that identification of who is high-risk constitutes ‘precision medicine’, even if treatment remains a population-average intervention. This is not precision medicine but ‘actuarial triage’.

Consider two patients each predicted to have an 80% risk of antipsychotic-induced metabolic syndrome, with similarly elevated baseline lipids and glucose. On paper, although they appear identical in risk profile, the model offers zero guidance on which modifiable risk factor to target – lipid management, glucose control or both? However, patients may respond very differently to the same intervention. Patient A might experience improvement with focused lipid management but patient B might see little change unless glucose is aggressively managed. This is the heterogeneity-of-treatment effect: intervention affects individuals differently. The deployed prediction model, however, is blind to this heterogeneity: it estimates P(Y = 1∣X), the probability of metabolic syndrome given current risk factors, but it does not estimate how much that probability will change for the individual if a specific intervention is applied: P(Y = 1∣do X). Faced with identical risk scores and no causal guidance, clinicians default to generic advice and standard protocols, i.e. the prediction alters urgency, not action. Many such interventions (eat healthily, exercise more, commence a metabolically safer antipsychotic) would be recommended regardless of whether predicted risk were 80 or 20%, creating the illusion of individualisation without changing clinical logic.

Unless explicitly modelled, ‘toggling’ lipid levels in a risk calculator to show that reducing lipid levels improves outcome, creates a false sense of informing an intervention. If a model has not learned the causal effect of lipid-lowering interventions, it cannot justify the claim that controlling lipids will improve outcomes for this patient. Apparent risk reductions reflect interface manipulation, not evidence of benefit.

This is the façade of precision: numerical risk (e.g. 80%) confers an aura of objectivity and tailoring, while clinical logic remains population-level and generic. Prediction models are not ‘objective tests’ and they encode the biases of their training data, and their probabilistic outputs, often poorly understood by clinicians, do not constitute causal guidance for individual care. ^{Reference Gigerenzer, Gaissmaier, Kurz-Milcke, Schwartz and Woloshin4,Reference Krishnadas5}

Prediction without causality can be harmful

These problems are not merely conceptual but are already evident in practice. A now-classic example comes from pneumonia risk prediction. In the 1990s, highly accurate models learned that patients with pneumonia and a history of asthma had lower risk of mortality compared with those without asthma. ^{Reference Caruana, Lou, Gehrke, Koch, Sturm and Elhadad6,Reference Cooper, Aliferis, Ambrosino, Aronis, Buchanan and Caruana7} This association was real but misleading. Asthmatic patients were routinely admitted to intensive care units, and the aggressive treatment they received reduced mortality. The model captured the association – pneumonia + asthma = lower risk – but not the causal mechanism. Deployed prospectively, such a model could have recommended against hospitalisation, paradoxically endangering high-risk patients.

An analogous problem arises in psychiatric risk tools built on routine clinical data. If high-risk patients in the training data received unmodelled interventions, such as crisis support, lithium or clozapine, or substance-misuse treatment, the model learns only that certain high-risk profiles are associated with better outcomes. Deployed prospectively, it may label similar new high-risk patients as ‘low risk’ and one may withhold the very interventions that produced those outcomes. ^{Reference Cockburn and Large8}

This is not a technical footnote, but a structural flaw: non-causal prediction models can encode and propagate dangerous treatment biases, systematically misclassifying those most likely to benefit from care. ^{Reference Krishnadas5}

Causal inference is not a minor technical detail

Seyedsalehi et al acknowledge that causal inference in observational studies relies on ‘strong and untestable assumptions’, but treat these as manageable nuisances. In routine psychiatric data, however, these assumptions are typically violated. Exchangeability is implausible when key drivers of treatment and outcome, such as insight, trauma history, personality traits and clinician preference, are unmeasured or crudely captured.

Non-adherence illustrates the problem: patients who discontinue antipsychotics often do so because of factors that independently worsen prognosis, e.g. substance use. Observing poor outcomes among non-adherent patients therefore does not establish that improving adherence will improve outcomes. Time-varying feedback among symptoms, treatment and adherence further compounds this. Framing these issues as tractable through ‘awareness’ substantially underestimates the depth of the causal challenge.

We recognise that fully specified counterfactual models are difficult to implement in routine psychiatry, where data are sparse and noisy. Progress will therefore be incremental, via hybrid models that incorporate key interventions and major confounders. Our point is about direction: precision psychiatry should aim to approximate individual treatment effects, even imperfectly, rather than stopping at ever more refined risk labels.

Finally, appeals to external validation offer false reassurance. Validation shows that a model reproduces associations in new data, not that its implicit causal assumptions are correct. ^{Reference Krishnadas, Leighton and Jones2} A model can validate extremely well if the same confounding structure persists in the validation cohort, encouraging clinicians to infer reasonably, but wrongly, that it supports sound action. In truth, population-level calibration and discrimination say nothing about whether changing a treatment decision based on the model will help or harm a specific patient.

Seyedsalehi et al call for nuance but end up defending a comfortable, low bar: models that rank risk and justify guideline adherence, rebranded as personalised medicine/precision psychiatry. Precision psychiatry should insist on leveraging causal structures established by randomised controlled trials to inform estimates of individual-level treatment effects, even where such estimates are necessarily approximate. Until then, we should be honest: current non-causal prediction models, at best, deliver risk stratification, not precision medicine, and their veneer of objectivity risks legitimising generic care as if it were tailored medicine. None of this implies that probabilistic risk stratification has no role in psychiatry. When used transparently to aid triage, non-causal models can be helpful providing they are not rebranded and presented as tools for personalised treatment or individual treatment choice. Precision psychiatry is about changing individual trajectories, not just forecasting them. Precision is meaningful to a patient only if their clinician can explain ‘why’ an outcome is likely and ‘what’ may help. Without this, probabilistic labels risk being ‘imposed’ rather than ‘shared’. Precision psychiatry can formalise and test the causal reasoning already present in careful, longitudinal clinical formulation – not replace it. Any modelling strategy that cannot credibly inform ‘what should we do, for this person, now?’ falls short of that standard.