# Cardiorespiratory Markers of Type 2 Diabetes: Machine Learning–Based Analysis

## TL;DR

This study uses machine learning to analyze heart and breathing patterns to distinguish people with and without type 2 diabetes, finding that certain cardiorespiratory metrics perform well.

## Contribution

The study introduces a systems-based framework combining HRV, FRF, and IR metrics with machine learning to better detect diabetes-related autonomic changes.

## Key findings

- Impulse response (IR) features showed strong standalone performance in distinguishing T2DM patients.
- Combining HRV and FRF metrics achieved the highest classification accuracy using SVM RBF under NM balancing.
- Systems-based approaches using multiple cardiorespiratory metrics may better capture diabetes-related autonomic differences than HRV alone.

## Abstract

The global prevalence of type 2 diabetes mellitus (T2DM) poses significant challenges due to its association with increased cardiovascular risk and complications like cardiovascular autonomic neuropathy. Measures derived from heart rate variability (HRV) and cardiorespiratory interactions quantified through frequency response function (FRF) and impulse response (IR) metrics reflect different aspects of autonomic regulation and may provide complementary physiological information relevant to diabetes-related autonomic alterations.

The study aimed to investigate whether these metrics, individually or in combination, provide useful physiological features for distinguishing individuals with and without T2DM using machine learning classifiers.

Electrocardiogram and respiratory signals from 2 PhysioNet datasets were used to derive 3 domains of autonomic and cardiorespiratory features: (1) spectral HRV indices reflecting overall variability; (2) FRF metrics characterizing frequency-specific respiratory-cardiac transfer properties; and (3) causal IR metrics capturing time-domain responsiveness to respiratory inputs. ML classifiers—logistic regression, support vector machine (SVM) with linear kernel, and SVM with radial basis function (SVM RBF) kernel—assessed the predictive value of individual and combined feature sets under NearMiss-1 (NM) undersampling and Synthetic Minority Oversampling Technique oversampling. This systems-based framework may capture subtle differences in respiratory-cardiac regulation associated with T2DM more effectively than HRV alone by reflecting integrated cardiorespiratory coupling.

Across classifiers and balancing strategies, IR features frequently produced comparatively strong standalone performance, suggesting that causal, time-domain cardiorespiratory dynamics capture informative physiological differences between groups. With logistic regression and NM, IR features achieved mean accuracy of 0.770 (SD 0.179), precision of 0.783 (SD 0.217), recall of 0.900 (SD 0.224), and F1-score of 0.798 (SD 0.140). While HRV metrics were the least informative standalone feature set, the combined HRV+FRF feature set under NM yielded the highest observed performance, with accuracy of 0.830 (SD 0.172), precision of 0.800 (SD 0.183), recall of 0.933 (SD 0.149), and F1-score of 0.853 (SD 0.145; SVM RBF). Under Synthetic Minority Oversampling Technique, HRV+IR showed the strongest observed combined performance, yielding accuracy of 0.700 (SD 0.128), precision of 0.783 (SD 0.217), recall of 0.683 (SD 0.207), and F1-score of 0.691 (SD 0.097) with SVM RBF, surpassing standalone IR in most metrics, though IR alone retained superior recall (0.950, SD 0.112) and F1-score (0.708, SD 0.038). These results reflect that performance depends on both feature domain and sampling strategy and that combining features capturing complementary physiological aspects of autonomic regulation may enhance discriminative ability.

HRV, FRF, and IR metrics each reflect distinct dimensions of autonomic and cardiorespiratory regulation. Systems-based approaches incorporating frequency-domain and causal dynamic features may offer richer characterization of diabetes-related regulatory differences than HRV alone. Although preliminary and limited by sample size, these findings highlight promising physiological feature domains and sampling strategies for future investigation. Larger datasets with well-defined autonomic phenotyping are needed to evaluate generalizability and determine clinical relevance.