Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Renaissance of RNNs in Streaming Clinical Time Series: Compact Recurrence Remains Competitive with Transformers

Ran Tong, Jiaqi Liu, Su Liu, Xin Hu, Lanruo Wang

TL;DR

The paper tackles streaming clinical time series under strict causality and limited data budgets by benchmarking a compact GRU-D RNN against a small Transformer on the MIT-BIH Arrhythmia dataset across two tasks: near-term tachycardia risk and one-step heart-rate forecasting. It employs calibration-aware evaluation with temperature scaling and bootstrap-based uncertainty, showing task-dependent results: recurrent encoders remain competitive for short-horizon risk scoring, while compact Transformers provide clearer gains for point forecasting. The study highlights the continued importance of calibration and evaluation under real-world deployment constraints and introduces a reproducible, strictly causal benchmark for bedside monitoring. These findings inform deployment strategies that balance latency, data availability, and performance depending on the clinical objective.

Abstract

We present a compact, strictly causal benchmark for streaming clinical time series on the MIT--BIH Arrhythmia Database using per-second heart rate. Two tasks are studied under record-level, non-overlapping splits: near-term tachycardia risk (next ten seconds) and one-step heart rate forecasting. We compare a GRU-D (RNN) and a Transformer under matched training budgets against strong non-learned baselines. Evaluation is calibration-aware for classification and proper for forecasting, with temperature scaling and grouped bootstrap confidence intervals. On MIT-BIH, GRU-D slightly surpasses the Transformer for tachycardia risk, while the Transformer clearly lowers forecasting error relative to GRU-D and persistence. Our results show that, in longitudinal monitoring, model choice is task-dependent: compact RNNs remain competitive for short-horizon risk scoring, whereas compact Transformers deliver clearer gains for point forecasting.

Renaissance of RNNs in Streaming Clinical Time Series: Compact Recurrence Remains Competitive with Transformers

TL;DR

The paper tackles streaming clinical time series under strict causality and limited data budgets by benchmarking a compact GRU-D RNN against a small Transformer on the MIT-BIH Arrhythmia dataset across two tasks: near-term tachycardia risk and one-step heart-rate forecasting. It employs calibration-aware evaluation with temperature scaling and bootstrap-based uncertainty, showing task-dependent results: recurrent encoders remain competitive for short-horizon risk scoring, while compact Transformers provide clearer gains for point forecasting. The study highlights the continued importance of calibration and evaluation under real-world deployment constraints and introduces a reproducible, strictly causal benchmark for bedside monitoring. These findings inform deployment strategies that balance latency, data availability, and performance depending on the clinical objective.

Abstract

We present a compact, strictly causal benchmark for streaming clinical time series on the MIT--BIH Arrhythmia Database using per-second heart rate. Two tasks are studied under record-level, non-overlapping splits: near-term tachycardia risk (next ten seconds) and one-step heart rate forecasting. We compare a GRU-D (RNN) and a Transformer under matched training budgets against strong non-learned baselines. Evaluation is calibration-aware for classification and proper for forecasting, with temperature scaling and grouped bootstrap confidence intervals. On MIT-BIH, GRU-D slightly surpasses the Transformer for tachycardia risk, while the Transformer clearly lowers forecasting error relative to GRU-D and persistence. Our results show that, in longitudinal monitoring, model choice is task-dependent: compact RNNs remain competitive for short-horizon risk scoring, whereas compact Transformers deliver clearer gains for point forecasting.
Paper Structure (4 sections, 14 equations, 1 figure, 2 tables)

This paper contains 4 sections, 14 equations, 1 figure, 2 tables.

Table of Contents

  1. Introduction and Related Work
  2. Methodology
  3. Experiments
  4. Conclusion and Future Work

Figures (1)

  • Figure 1: Tachycardia classification on MIT--BIH (mean$\pm$std over three seeds). Higher is better for AUROC/AUPRC/F1; lower is better for Brier/ECE. GRU-D shows higher AUROC/AUPRC and lower Brier than the compact Transformer, while F1 at the validation-selected threshold is similar across models. ECE remains nontrivial for both, underscoring the need for calibration.