Can a Sensor Sense Depression?

Motivation

Your router, your voice, your daily rhythm: all of them leak signals about how you are doing. Can those signals become biomarkers, trusted the way a blood test is trusted, but read from everyday sensors instead of a lab [4], [5]?

The hard part is trust, not the answer. Feed data to an AI model and you get a “depression score.” But a score you cannot explain is not a biomarker, it is a guess. A biomarker must be measurable, reproducible, and tied to a clinical meaning, here the DSM-5 criteria [1], which today are assessed mostly through subjective self-reports [3]. Even summing symptoms into one score is contested [2]. Black-box models, large language models included [8], predict without defining or validating a marker, and usually send private data to the cloud to do it.

The group builds markers the honest way: behaviour from home-router traffic (CareNet), speech acoustics (IHearYou) [6], [7], and a multimodal, context-aware framework (our CML framework). Each maps features transparently to DSM-5 indicators and runs on the device.

Problem statement

Sensor pipelines today predict a depression label; they do not define a biomarker. They report accuracy, but not whether a signal is measurable, reproducible, and tied to a specific DSM-5 indicator, nor whether it holds up across people, devices, and contexts. That gap is the difference between a demo and a marker a clinician could trust.

This thesis fills it: you will define what makes a sensor-derived signal a valid, interpretable digital biomarker, and build a framework that tests candidate markers for reproducibility, robustness to confounders, and clinical grounding, not just predictive accuracy.

Research questions

1. What makes a sensor-derived signal a valid digital biomarker of a DSM-5 depressive indicator, rather than a confounded correlation?
2. Can we define and validate interpretable digital biomarkers that are reproducible and robust to confounders such as age, sex, recording device, and context?
3. On validity, interpretability, and privacy, and not only accuracy, how do such biomarkers compare to black-box predictors [8], [9]?

Suggested methodology

1. Pick a sensing modality that appeals to you (speech, network, or another) and extract candidate markers, building on our transparent feature-to-DSM-5 mappings.
2. Formalise an interpretability-first definition of a digital biomarker: measurable, reproducible, and explicitly linked to a clinical indicator.
3. Build a small validation framework that tests reproducibility, robustness to confounders, and generalisation to a second dataset (e.g. DAIC-WOZ [10] for speech).
4. Compare the resulting markers, on trust and privacy and not only accuracy, against a black-box baseline.
5. Deliver a working on-device prototype and a clear, evidence-based answer to “what makes a digital biomarker trustworthy?”

You will take machine learning through a real health problem, end to end. You need machine-learning basics and Python, plus curiosity about health and interpretability. No sensing or signal-processing background is required, we bring that side.

Previous theses in this line

• At the Edge of Empathy: Mental-Health Awareness Through Home-Router Network Traffic Insights, Stephan Nef (CareNet).
• Audio-centered Approach for Building a Multimodal Predictive AI Agent to Detect Depressive Behaviors, Jonas Länzlinger (IHearYou).
• Explainable Multimodal-based Depression Awareness at the Edge, Tibor Haller (CML).

References

[1] American Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders: DSM-5-TR, 5th ed. Washington, DC, 2022.

[2] E. I. Fried and R. M. Nesse. “Depression sum-scores don’t add up: why analyzing specific depression symptoms is essential.” BMC Medicine, 13(1):72, 2015.

[3] K. Kroenke, R. L. Spitzer, and J. B. Williams. “The PHQ-9: validity of a brief depression severity measure.” Journal of General Internal Medicine, 16(9):606-613, 2001.

[4] S. Saeb, M. Zhang, C. J. Karr, et al. “Mobile phone sensor correlates of depressive symptom severity in daily-life behavior: an exploratory study.” Journal of Medical Internet Research, 17(7), 2015.

[5] R. Wang, F. Chen, Z. Chen, et al. “StudentLife: assessing mental health, academic performance and behavioral trends of college students using smartphones.” ACM UbiComp, 2014.

[6] N. Cummins, S. Scherer, J. Krajewski, et al. “A review of depression and suicide risk assessment using speech analysis.” Speech Communication, 71:10-49, 2015.

[7] D. M. Low, K. H. Bentley, and S. S. Ghosh. “Automated assessment of psychiatric disorders using speech: a systematic review.” Laryngoscope Investigative Otolaryngology, 5(1):96-116, 2020.

[8] A.-M. Bucur. “Leveraging LLM-generated data for detecting depression symptoms on social media.” CLEF, Springer, 2024, pp. 193-204.

[9] L.-V. Herm, K. Heinrich, J. Wanner, and C. Janiesch. “Stop ordering machine learning algorithms by their explainability! A user-centered investigation of performance and explainability.” International Journal of Information Management, 69:102538, 2023.

[10] USC Institute for Creative Technologies. “DAIC-WOZ database.” https://dcapswoz.ict.usc.edu/