Lecture: A Guided Safari Through the Time-Series Modelling Zoo for Health

Healthcare and life-science data mostly arrive as sequences: vital-sign streams, ICU waveforms, wearable data and multimodal signals such as speech and movement, longitudinal lab data, and patient journeys in electronic health records. This lecture is a guided “safari” through the time-series modelling zoo—designed to choose the right model family sensibly for real health problems. We start in the “classic enclosure” with alignment and similarity methods, including Dynamic Time Warping (DTW) and common variants, followed by probabilistic sequence models such as Hidden Markov and Graphical Models (HMMs and GMMs). We then visit neural sequence learning with RNNs, LSTMs, and modern extensions such as xLSTMs. Next, we move to the attention habitat—Transformers for continuous signals and patient trajectories—covering key design choices for temporal encoding, tokenization/patching, and multimodal fusion. Finally, we explore the “new arrivals”: state-space models and efficient long-context sequence modelling for high-frequency signals and long patient histories. We will be emphasizing practical challenges that dominate healthcare data: irregular sampling, missingness, limited labels, and likes. Rather than treating this as a mere model parade, the session is anchored in recurring healthcare tasks—early warning and deterioration prediction, forecasting, monitoring, and personalized prediction. You shall leave with a coherent map of the landscape and a practical decision framework for selecting and validating time-series models in health and life-science applications.