Poster L29 - Rapid Classification of Vascular Aging from Continuous Photoplethysmography Monitoring of Trauma Patients

Introduction: Physiological fitness is a critical predictor of prognosis following injury. It guides expectations of clinical severity, treatment effects and overall outcomes. Aging is an endogenous mediator of physiological fitness changes over a lifetime, bringing with it epigenetic modifications, decreases in pulmonary function, and even increased arterial wall stiffness. Research has shown that changes in arterial wall stiffness with age correspond to morphological changes in photoplethysmography (PPG) waveforms, albeit these findings have only been substantially tested in healthy populations. We hypothesize that age-mediated changes to PPG morphology can allow for rapid classification of physiological age in a general trauma cohort, and furthermore that models performing such tasks are minimally influenced by autonomic dysfunction.

Materials and

Methods: A total of 20,008 patients were retrospectively identified from a population of 29,329 consecutive adult trauma admissions to a level one Shock Trauma Center in Baltimore, MD between 2009 and 2018. Patients were excluded for having an age less than 18 or greater than 99. The first 15 minutes of a patient’s continuous vital signs (PPG) in the Trauma Resuscitation Unit were used to generate two sets of predictive features. These were either PPG-derived heart rate variability (HRV) features or 11 features related to waveform morphology of the dicrotic notch. The total patient dataset was sequestered into a training/validation set, which was comprised of the first 2/3 of the dataset by earliest admissions. The remaining 1/3 of the dataset was used for model testing. The outcome metric was binary and identified as those with young age (≤ 38 years old) and those with old age (≥ 60 years old). These categories were chosen based on previous studies identifying healthy vascular aging populations using vital signs. Preprocessing was performed to scale features, tune hyperparameters, and perform a minimized Redundance Maximized Relevance feature selection. A 10x10 cross-validated gradient boosting tree model was developed on the training/validation dataset, with the testing set blinded to all preprocessing steps. The model utilized binary classification to predict whether the patient belonged to the old or young age category. Model performance was scored via the area under the receiver operator characteristic curve (AUROC). Shapley Additive Explanations Analysis (SHAP) identified which features contributed most to the test-set performance.

Results, Conclusions, and Discussions: The baseline characteristics of patients selected were age: 40.0 years old, 30.9% women, 72.5% blunt injury with an initial GCS of 15. Dicrotic notch features alone achieved a validation AUROC of 0.81 ± 0.01 and a testing AUROC of 0.77. The inclusion of all PPG features in the model yielded a validation performance of 0.92 ± 0.01 and a testing performance of 0.89. SHAP analysis of the PPG model indicated that the most important features were either a) related to notch morphology or b) non-linear dynamics HRV variables derived from PPG. These results were stable when filtering the dataset and analyzing model performance on a traumatic brain injured population, a surrogate population with an expected increased incidence of autonomic dysfunction.

Our study shows that photoplethysmography signals can rapidly and reliably classify patient age with high sensitivity and specificity in a sick-patient population. Furthermore, the model performance regardless of disease type indicates that the critical information may lie in the waveform morphology and not peak-to-peak information modulated by the autonomic nervous system. Our future plans aim to develop machine learning-generated “physiologic age” determinations to enhance the predictions of severity-compensation levels, clinical responses to treatment, and longer-term prognostics.

Acknowledgements (Optional):