Uncertainty-aware learning from sparse, unlabelled, and out-of-distribution time series

Nowadays, machine learning is increasingly popular in the analysis of healthcare time series, as it can support improved diagnostics, personalised monitoring, and effective performance tracking. The ever-increasing availability of datasets from wearable sensors, mobile devices, and continuous monitoring technologies has created new opportunities for real-world applications, as they can offer a rich and continuous picture of an individual's health and fitness in everyday environments as well as in clinical or supervised healthcare settings. Despite the potential that machine learning models offer for healthcare time series, they still face notable challenges. Sensor-based datasets are frequently sparse (with missing values) if acquired outside controlled environments, while a substantial proportion remain unlabelled or only partially labelled. Further, models often exhibit limited generalisability under out-of-distribution conditions, which arise due to heterogeneity across hospitals, sensors, and patient populations. Such factors limit the use of real-world time series in deep learning applications, and are consequential in healthcare as unreliable models may not only yield poor performance but also produce overconfident metrics. In this context, this thesis moves beyond the narrow settings of conventional models through uncertainty-aware learning. Effectively incorporating uncertainty enhances training by addressing key challenges, as it informs the handling of missing data, prioritises the most informative samples for annotation, and detects distribution shifts in out-of-distribution data to guide fine-tuning. By leveraging uncertainty, this thesis makes models more data-efficient, adaptive, and generalisable across diverse healthcare use cases. The first contribution focuses on incorporating uncertainty-aware sequence-to-sequence predictions on sparse time series. These are often sparsely-sampled due to missed recordings, device malfunctions, or diverse sensing conditions. By enhancing uncertainty estimation in evidential deep learning and introducing metrics for assessing uncertainty estimations on sequence-to-sequence predictions, the proposed methodology provides a reliable way to design sequence-to-sequence prediction models for real-world use cases. Further, this thesis proceeds with a contribution addressing the bottleneck of unlabelled time series by automating machine learning workflows in a way that integrates human expertise and automated model refinement. Recognising the costly nature of labelling biosignals and the limited technical expertise of medical experts in model optimisation, the proposed approach facilitates the selection of highly-informative samples for annotation, dynamically tunes models during training, and maximises the use of unlabelled data. This method continuously refines the model with expanding data and human input, outperforms baselines and state-of-the-art methods, and reduces reliance on manual annotation and parameter tuning, thereby maximising the information gained through each annotation step. Finally, this thesis also focuses on enhancing model adaptability under distribution shifts in healthcare time series, where datasets collected in one context often differ from those encountered in deployment. By leveraging uncertainty to recognise, quantify, and adapt to diverse and overlapping shifts, the proposed method improves the fine-tuning of pre-trained models under out-of-distribution conditions. This allows models to better handle multiple types and levels of distribution shifts, ensuring more reliable and effective adaptation, while avoiding overfitting and inefficient use of computational resources. Overall, by addressing the limitations imposed by sparse data, unlabelled data, and distribution shifts, this thesis contributes to the design of more generalisable, data-efficient, and adaptive algorithms for healthcare time series analysis.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Vavaroutas, Sotirios

Advisor dc:contributor.advisor

• Mascolo, Cecilia

• Artificial Intelligence
• Machine Learning
• Uncertainty-Aware Learning
• Sparse Data
• Unlabelled Data
• Out-of-Distribution Data
• Time Series

Licence

• https://www.repository.cam.ac.uk/bitstreams/98d1fc07-efcb-4175-acfc-88330924f467/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.129978

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/402673