On the effectiveness of smartphone IMU sensors and Deep Learning in the detection of cardiorespiratory conditions

TL;DR

The paper presents a low-cost, noninvasive screening approach for cardiorespiratory diseases using commodity smartphone IMUs and deep learning. By collecting breathing kinematics from five body regions and segmenting breathing cycles, the authors train a bidirectional LSTM encoder to produce cycle-aware embeddings that feed a dense classifier, validated with Leave-one-out cross-validation and Bayesian hyperparameter optimization. The Bi-LSTM_128,2 configuration achieves robust performance (test metrics around 0.81–0.82 for sensitivity and specificity, 80.2% accuracy) and generalizes to skewed healthy-only distributions, suggesting utility for at-home population screening and pandemic response. The study demonstrates the potential of widely available smartphones for remote, scalable cardiorespiratory screening and disease progression monitoring, while acknowledging limitations such as dataset size, diversity, and sensor variability that warrant further work.

Abstract

This research introduces an innovative method for the early screening of cardiorespiratory diseases based on an acquisition protocol, which leverages commodity smartphone's Inertial Measurement Units (IMUs) and deep learning techniques. We collected, in a clinical setting, a dataset featuring recordings of breathing kinematics obtained by accelerometer and gyroscope readings from five distinct body regions. We propose an end-to-end deep learning pipeline for early cardiorespiratory disease screening, incorporating a preprocessing step segmenting the data into individual breathing cycles, and a recurrent bidirectional module capturing features from diverse body regions. We employed Leave-one-out-cross-validation with Bayesian optimization for hyperparameter tuning and model selection. The experimental results consistently demonstrated the superior performance of a bidirectional Long-Short Term Memory (Bi-LSTM) as a feature encoder architecture, yielding an average sensitivity of $0.81 \pm 0.02$, specificity of $0.82 \pm 0.05$, F1 score of $0.81 \pm 0.02$, and accuracy of $80.2\% \pm 3.9$ across diverse seed variations. We also assessed generalization capabilities on a skewed distribution, comprising exclusively healthy patients not used in training, revealing a true negative rate of $74.8 \% \pm 4.5$. The sustained accuracy of predictions over time during breathing cycles within a single patient underscores the efficacy of the preprocessing strategy, highlighting the model's ability to discern significant patterns throughout distinct phases of the respiratory cycle. This investigation underscores the potential usefulness of widely available smartphones as devices for timely cardiorespiratory disease screening in the general population, in at-home settings, offering crucial assistance to public health efforts (especially during a pandemic outbreaks, such as the recent COVID-19).

Figures (8)

• Figure 1: Acquisition protocol. (Left) The five targeted positions for smartphone placement (three on the chest, and two on the abdomen). (Middle) The patient is laying on a bed with a smartphone placed in the middle chest position (M1), illustrating how the smartphone records acceleration and angular velocity through the dedicated app and safely transmits the data to a server (Right).
• Figure 2: Joint plot in the first column showing the distribution of height and weight characteristics among healthy and preoperative patients. The remaining columns feature different boxplots illustrating the distribution of age, height, and weight across various health conditions.
• Figure 3: Time series plots of normalized and filtered accelerometer and gyroscope signals from five different recording locations, comparing healthy control and positive class patients. Signals are displayed for a duration of 20 seconds (1000 timesteps) for plotting convenience.
• Figure 4: (Top) Low-pass signal filtering on gyroscope Y-axis and peaks identification. Upfacing and downfacing markers denote inhalation and exhalation phases respectively. (Bottom) Signal windowing on original raw data. The same windowing is also applied to time series data coming from the accelerometer.
• Figure 5: Visual workflow of the predictive model. Each processed input signal representing accelerometer and gyroscope data is processed through a bidirectional LSTM encoder. The concatenated embeddings are fed to fully connected layers providing a prediction regarding the presence of a cardiorespiratory disease.