Psychosis Prognosis Prediction: integrating human perspectives and artificial intelligence

Psychotic disorders have highly variable outcomes—while some individuals recover fully, others continue to struggle with severe symptoms and functional impairment. Currently, it remains difficult to predict individual prognosis due to variability in predictors, inconsistent outcome definitions, and methodological heterogeneity. Machine learning (ML) holds promise for enhancing individualized prognosis prediction, but clinical integration requires validation, transparency, and alignment with stakeholder needs. This thesis aimed to combine public and patient involvement (PPI) with ML-driven approaches to improve individual outcome predictions and support more personalized care.

Part 1 of the thesis explored predictors and outcome measures in psychotic disorders from multiple stakeholder perspectives. In Chapter 2, we conducted a systematic review and meta-analysis of 178 prospective studies on short-term (≤1 year) outcomes in schizophrenia spectrum disorders (SSD). We found that lower chances of symptomatic remission were associated with male gender, longer duration of untreated psychosis, more severe symptoms, worse global functioning, a higher number of previous hospitalizations, and poor treatment adherence. Risk of readmission was higher with more prior admissions, and poor baseline functioning reduced the likelihood of functional improvement. For other commonly cited predictors, like age at onset or depressive symptoms, little to no consistent evidence was found. We attribute discrepancies with prior findings to methodological limitations in earlier research and emphasize the need for open-access data and analysis scripts to facilitate replication and meta-research.

In Chapter 3, we examined differing stakeholder priorities regarding outcomes in psychosis. An online survey was completed by 106 service users (SUs), 51 informal caregivers (ICs), and 69 healthcare professionals (HCPs). While broad agreement existed, differences emerged in emphasis: SUs and ICs valued clinical, functional, and personal recovery equally, whereas HCPs prioritized clinical recovery. Functional recovery was viewed differently as well—SUs emphasized day-to-day activities (e.g., chores), while ICs focused on underlying skills like planning and organizing. The healthcare system itself was seen as both facilitator (e.g., therapeutic relationships) and barrier (e.g., fragmented care), with ICs emphasizing its importance more than HCPs. These findings highlight the need for stakeholder dialogue to align priorities in recovery-oriented care.

Part 2 investigated the opportunities and barriers for applying ML models to prognosis prediction in psychosis. Despite growing research, no ML model has yet been adopted in clinical practice. In Chapter 4, we assessed the performance of an ML model from the Psychosis Prognosis Predictor project. Its predictive accuracy was modest but comparable to that of psychiatrists. Though not superior, the model may serve as a useful second opinion in complex cases. Psychiatrists stressed the need for greater model transparency and interpretability.

To that end, Chapter 5 explored counterfactual model explanations—hypothetical scenarios showing how outcome predictions change with altered inputs. Using this technique, we demonstrated the differing impact of psychiatric comorbidities on remission prediction across patient subgroups. Such approaches not only increase transparency but may also empower patients to modify risk factors (e.g., cannabis use, exercise) and actively engage in improving their prognosis.