AAI_2025_Capstone_Chronicles_Combined

7

Empirical evidence supports the effectiveness of CNN – LSTM models for ICU mortality

prediction. For example, (Yeh et al., 2024). achieved an area under the receiver operating characteristic

curve (AUROC) of 87.8 % and an area under the precision-recall curve (AUPRC) of 53.5 % on the

MIMIC-III dataset using CNN – LSTM, outperforming standalone CNN and LSTM models. Similarly,

(Khan, 2019) demonstrated that CNN – LSTM models capture both semantic and temporal relationships in

electronic health record (EHR) time series, resulting in higher predictive accuracy compared to classical

statistical methods. These findings, along with the architec ture’s ability to integrate localized event

detection with long-term trend modeling, underscore CNN –LSTM’s suitability for ICU mortality

prediction within clinically actionable timeframes.

Transformers

Lastly explored transformer models, originally developed for natural language processing, have

become increasingly relevant for time-series prediction in healthcare. Their self-attention mechanisms

enable efficient modeling of long-range dependencies while accommodating irregular sampling and

missing data, challenges that are especially prevalent in ICU datasets such as PhysioNet (msafi04, 2021).

Unlike sequential models like LSTMs, Transformers process entire sequences in parallel, making them

well-suited for capturing complex temporal interactions among physiological variables. Their modular

architecture also facilitates interpretability and supports probabilistic outputs, both of which are critical

for high-stakes clinical decision-making.

Recent studies have demonstrated the versatility of Transformer-based models across a range of

healthcare applications. BEHRT (Li et al., 2020) adapted the Transformer framework for disease

trajectory modeling using electronic health records, while ClinicalBERT (Alsentzer et al., 2019) was fine

tuned for clinical note classification. In ICU-specific tasks, Transformer models have outperformed

153