2026 ECNP Congress

EP01-1033

Latent behavioural states obtained from passive digital phenotyping and their clinical characterisation

Ao Zhang — N. Cummins, Y. Zhang, H. Sankesara, F. Matcham, S. Siddi, F. Lamers, A.A. Folarin, G. Breen, B.W.J.H. Penninx, T. Wykes, R.J.B. Dobson, V.A. Narayan, M. Hotopf, V. Srini · King’s College London, Institute of Psychiatry- Psychology and Neuroscience, London, United Kingdom

2026 ECNP Congress

Background: Smartphone and wearable sensing can capture mobility, sleep, geolocation, and device-regularity behaviours relevant to depression [1], [2], [3], [4]. Prior work in major depressive disorder (MDD) has linked home stay, wearable-derived sleep features, and smartphone-derived nearby-device-count features to Patient Health Questionnaire-8 (PHQ-8) symptom severity [1], [3], [4]. Recent analyses suggest that approximately eight days of data are often needed for stable feature estimation, and that depression-related behavioural signals differ across individuals and time [2]. However, most studies summarise behaviour with hand-crafted averages across fixed windows, features in isolation, or mean-pooled representations. These approaches can detect associations but may obscure within-person transitions between recurring behavioural modes. This matters in MDD, where fluctuating symptomatology motivates objective tracking over time.

Aims: This study aimed to (i) learn interpretable latent states from passive sensing; (ii) characterise within-participant state evolution; and (iii) test whether temporal state structure supports candidate MDD subtype clustering beyond average behaviour alone. The approach combines convolutional autoencoding with state-based temporal summaries, preserving recurrence, persistence, instability, and transition information that mean-pooled profiles discard.

Methods: Passive data from the Remote Assessment of Disease and Relapse - MDD (RADAR-MDD) cohort [5] were represented as participant-day 24 × 6 matrices comprising hourly phone usage, communication app usage, social app usage, time at primary location, time away from primary location, and travelling. After quality control, a convolutional variational autoencoder with a four-layer encoder learned a six-dimensional latent embedding for each participant-day; pooled embeddings were clustered to identify recurrent behavioural states using candidate approaches including MiniBatchKMeans, Gaussian mixture models, and HDBSCAN, with participant-balanced resampling to assess robustness and reduce leakage from random day-level splits. Participant-level temporal summaries included state occupancy, transition frequencies, dwell time, switching entropy, latent volatility, and hidden Markov model-derived transition features. These features were compared with mean-pooled embeddings and aligned to 7-, 14-, and 21-day windows preceding PHQ-8 and GAD-7 assessments. Associations with symptom scores, severity bands, future worsening, and MDD status were examined using grouped participant-level splits and mixed-effects models.

Results: The latent-state pipeline represents participant-days as structured passive-sensing profiles and derives day-level embeddings for recurrent-state and transition analyses. This supports clinically interpretable summaries that can be reviewed alongside established symptom measures. At participant level, it preserves state occupancy, persistence, switching entropy, dwell time, and latent volatility, enabling comparison with mean-pooled embeddings and PHQ-8/GAD-7-aligned windows. Clustering these temporal summaries provides a candidate digital subtype framework for examining how recurrence, persistence, and instability relate to depressive symptom severity and MDD status.

Conclusions: A latent behavioural state framework bridges passive sensing and candidate MDD subtype structure by preserving recurrence, persistence, switching, and instability. This state-based representation moves digital phenotyping beyond static average behaviour and provides a basis for stratified analyses of depressive symptom burden, clinical status, and subsequent clinical validation.

References

  1. Laiou, P., Kaliukhovich, D.A., Folarin, A.A., Ranjan, Y., Rashid, Z., Conde, P., Stewart, C., Sun, S., Zhang, Y., Matcham, F., Ivan, A., Lavelle, G., Siddi, S., Lamers, F., Penninx, B.W.J.H., Haro, J.M., Annas, P., Cummins, N., Vairavan, S., Manyakov, N.V., Narayan, V.A., Dobson, R.J.B., Hotopf, M., RADAR-CNS, 2022. The association between home stay and symptom severity in major depressive disorder: preliminary findings from a multicenter observational study using geolocation data from smartphones. JMIR Mhealth Uhealth 10, e28095. https://doi.org/10.2196/28095.
  2. Sun, S., Folarin, A.A., Zhang, Y., Cummins, N., Garcia-Dias, R., Stewart, C., Ranjan, Y., Rashid, Z., Conde, P., Laiou, P., Sankesara, H., Matcham, F., Leightley, D., White, K.M., Oetzmann, C., Ivan, A., Lamers, F., Siddi, S., Simblett, S., Nica, R., Rintala, A., Mohr, D.C., Myin-Germeys, I., Wykes, T., Haro, J.M., Penninx, B.W.J.H., Vairavan, S., Narayan, V.A., Annas, P., Hotopf, M., Dobson, R.J.B., RADAR-CNS Consortium, 2023. Challenges in using mHealth data from smartphones and wearable devices to predict depression symptom severity: retrospective analysis. J Med Internet Res 25, e45233. https://doi.org/10.2196/45233.
  3. Zhang, Y., Folarin, A.A., Sun, S., Cummins, N., Bendayan, R., Ranjan, Y., Rashid, Z., Conde, P., Stewart, C., Laiou, P., Matcham, F., White, K.M., Lamers, F., Siddi, S., Simblett, S., Myin-Germeys, I., Rintala, A., Wykes, T., Haro, J.M., Penninx, B.W.J.H., Narayan, V.A., Hotopf, M., Dobson, R.J.B., RADAR-CNS Consortium, 2021. Relationship between major depression symptom severity and sleep collected using a wristband wearable device: multicenter longitudinal observational study. JMIR Mhealth Uhealth 9, e24604. https://doi.org/10.2196/24604.
  4. Zhang, Y., Folarin, A.A., Sun, S., Cummins, N., Ranjan, Y., Rashid, Z., Conde, P., Stewart, C., Laiou, P., Matcham, F., Oetzmann, C., Lamers, F., Siddi, S., Simblett, S., Rintala, A., Myin-Germeys, I., Wykes, T., Haro, J.M., Penninx, B.W.J.H., Narayan, V.A., Annas, P., Hotopf, M., Dobson, R.J.B., RADAR-CNS Consortium, 2021. Predicting depressive symptom severity through individuals’ nearby Bluetooth device count data collected by mobile phones: preliminary longitudinal study. JMIR Mhealth Uhealth 9, e29840. https://doi.org/10.2196/29840.
  5. Matcham, F., Barattieri di San Pietro, C., Bulgari, V., de Girolamo, G., Dobson, R., Eriksson, H., Folarin, A.A., Haro, J.M., Kerz, M., Lamers, F., Li, Q., Manyakov, N.V., Mohr, D.C., Myin-Germeys, I., Narayan, V., Penninx, B.W.J.H., Ranjan, Y., Rashid, Z., Rintala, A., Siddi, S., Simblett, S.K., Wykes, T., Hotopf, M., RADAR-CNS Consortium, 2019. Remote assessment of disease and relapse in major depressive disorder (RADAR-MDD): a multi-centre prospective cohort study protocol. BMC Psychiatry 19, 72. https://doi.org/10.1186/s12888-019-2049-z.