Digital phenotyping signatures of impending psychotic relapse: a machine learning approach using smartphone-derived behavioural data

Introduction

Most psychiatric services still detect relapse in schizophrenia spectrum disorders the same way they always have: by waiting for patients or carers to report something has gone wrong. The method is subjective, episodic, and reactive by design. Digital phenotyping, which captures behaviour passively through smartphone sensors, offers a different way of working. The data is objective, ecologically valid, and continuous. The theoretical case has been around for over a decade. What has changed in the last few years is that the empirical case has caught up, with several independent cohorts and platforms now reporting convergent findings.

Aims

To synthesise the evidence on behavioural digital signatures preceding psychotic relapse, evaluate the machine learning approaches applied to passively sensed smartphone data, and assess what clinical translation actually requires — methodologically, clinically, and from a regulatory standpoint.

Methods

Systematic narrative review of peer-reviewed publications from 2015 to 2025 on digital phenotyping in schizophrenia spectrum disorders. Passive sensing modalities reviewed: GPS-derived mobility indices, accelerometry, communication metadata, screen-state sleep proxies, and voice and speech features. Machine learning approaches assessed: random forests, long short-term memory networks, federated learning architectures, and individual-specific anomaly detection. Major research platforms examined: CrossCheck, mindLAMP, RADAR-CNS, and Beiwe.

Results

A handful of behavioural signatures have replicated across independent samples. Sleep fragmentation, derived from smartphone screen-state proxies, predicts prodromal symptom escalation at AUC 0.78. Geographic restriction — increased home-time and decreased radius of gyration from GPS traces — is a consistent precursor. Outgoing communication frequency drops and response latency lengthens. Speech coherence degrades, with reduced semantic density and prolonged pause-to-speech ratios detectable from prompted samples collected within dedicated applications. Random forest classifiers applied to passively sensed data have reached sensitivity of 83% and specificity of 76% for relapse prediction. Bidirectional long short-term memory networks applied to combined GPS and accelerometry achieve AUC 0.84 on seven-day-ahead prediction horizons. Federated learning approaches match centralised model performance with much lower data privacy exposure, which matters directly for European regulatory compliance.

The most important unresolved problem is not technical. It is clinical. No published study has specified, tested, or evaluated an explicit escalation protocol — the decision rule that converts a digital alert into a clinical action. Without that rule, sensitivity and AUC measure whether a signal exists, not whether acting on it changes anything for the patient. Under the EU AI Act, relapse prediction systems will probably qualify as high-risk AI; under the Medical Devices Regulation, Class IIa or IIb classification is the likely outcome. Neither framework has yet been applied prospectively to a deployed digital phenotyping system in psychiatry.

Conclusions

The behavioural signatures of psychotic relapse are real, measurable, and replicable. Whether acting on them improves outcomes for patients is a separate question, and one that has not yet been answered in a randomised trial powered for clinical endpoints. What the field needs now is exactly that trial, with an explicit escalation protocol and a regulatory-compliant data architecture.

References

References

1. Barnett I. et al., 2018. Relapse prediction in schizophrenia through digital phenotyping: a pilot study. Neuropsychopharmacology 43(8), 1660–1666.
2. Bedi G. et al., 2015. Automated analysis of free speech predicts psychosis onset in high-risk youths. NPJ Schizophrenia 1, 15030.
3. Millgate E. et al., 2023. Relapse prediction in schizophrenia with smartphone digital phenotyping during COVID-19. Schizophrenia 9, 43.
4. Onnela JP., Rauch SL., 2016. Harnessing smartphone-based digital phenotyping. Neuropsychopharmacology 41(7), 1691–1696.
5. Ranjan Y. et al., 2019. RADAR-Base: open source mobile health platform. JMIR Mhealth Uhealth 7(8), e11734.