The Apple Research team has released a study examining the ability of AI models to estimate heart rates from stethoscope audio recordings, even though they were not specifically designed for this purpose. The results are encouraging, showing that AI can accurately estimate heart rates from recordings of heart sounds, referred to as phonocardiograms.
This research tested six widely-used foundation models that were trained on audio or speech to assess their heart rate estimation capabilities. Interestingly, these models showed performance levels similar to those of traditional methods that depend on handcrafted audio features. Notably, Apple’s proprietary model, a version of CLAP (Contrastive Language-Audio Pretraining), surpassed the baseline, demonstrating the most effective performance in various assessments.
The evaluation process used a publicly accessible dataset comprising over 20 hours of labeled hospital heart sounds. Apple segmented these recordings into brief 5-second clips, resulting in around 23,000 heart sound snippets which were utilized to train a neural network for heart rate classification.
An important finding from the study indicated that larger models did not necessarily provide superior outcomes. In fact, the deeper layers of these models frequently contained less useful information for cardiorespiratory data, suggesting that representations from shallower or mid-layers were more effective.
The study highlights the importance of merging traditional signal processing methods with cutting-edge AI to improve the reliability of heart rate estimations. The researchers intend to enhance these models for healthcare applications, create lighter versions suitable for low-power devices, and investigate other physiological sounds.
While the study does not assert clinical claims, it underscores the potential for incorporating these models into Apple devices, including iPhones, Apple Watches, and AirPods, which could leverage in-ear microphones for monitoring heart rates.
For additional information, the complete paper is accessible on arXiv.
