I work with wearable sensor data to study acute health events in older adults. My work covers data preparation, model comparisons and interpretability. First-author paper with Prof. Krzysztof Grudzień at AIME 2026; further modelling work continues.
Working with the data
I processed 72.6 million measurements from 13,317 telecare patients. Filtering missing data left a cohort of 8,267 patients for later modelling work. I engineered physiological and circadian features, compared models and examined how thresholds would affect the number of alarms.
From the paper to the next study
Our AIME 2026 paper covers the earlier unsupervised anomaly-detection study. I’m now working on an extended manuscript with more data, shorter monitoring windows and more personalized detection methods. This remains research; the later modelling experiments are separate from the published study.
Two stages of the research
The AIME paper asked whether unusual patterns in longitudinal wearable data could help identify acute health events. Later work widened the modelling choices to supervised sequences, circadian features, tree models and combinations of these approaches. The diagram keeps those branches separate because they answer related questions with different experiments.
Looking beyond the score
A model score is only part of the question. I also look at which features influence predictions and how a threshold changes the number of potential alarms. The source patients, filtered cohort and individual experiments are different groups, so their counts need to stay attached to the relevant analysis. The public repository contains code and documentation, while reproducing the study requires access to the underlying data and settings.
