Matching Inference Depth to the Heartbeat

A faster heart rate leaves less time to classify the next ECG window. At North Carolina State University, I co-developed heart-rate-aware Adaptive Model Selection and a parameter-shared Anytime CNN, allowing inference depth to change with the available time budget.

The network combines ECG waveform features with inter-beat timing and exposes three early exits in one shared backbone. This gives the controller lightweight, intermediate, and full-depth inference paths without loading separate models. I also corrected preprocessing and label alignment, then benchmarked the system on Raspberry Pi 4.

Accuracy Within a Real-Time Budget

In the reported two-cycle evaluation, adaptive inference reached 91.5% accuracy and 90.6% F1 at 1.33 ms mean latency, with no deadline misses among 1,000 evaluated samples.

Two-Cycle Configuration Accuracy F1 Mean Latency Deadline Misses
AMS + Anytime CNN 91.5% 90.6% 1.33 ms 0 / 1,000
CNN-LSTM baseline 87.3% 87.6% 3.33 ms 1,000 / 1,000

Against the strongest literature-based baseline in this comparison, the adaptive configuration improved accuracy by 4.2 percentage points while reducing mean latency by approximately 60%. The result made the accuracy–latency trade-off an explicit runtime decision.

This work was published at IEEE RTCSA 2025; I am the second author of Adaptive Model Selection for Real-Time Heart Disease Detection on Embedded Systems.