Exploring explainable machine learning techniques to aid dysphagia risk identification: A feasibility study

Abstract

Background

Machine learning offers opportunities to identify complex risk patterns in large data sets. We explored the methodological feasibility, and proof of concept, of applying machine learning techniques to dysphagia (swallowing difficulty) risk identification for adult patients who required endotracheal intubation within an intensive care unit (ICU).

Aim

The aim of this study was to explore the methodological feasibility and proof of concept of developing machine learning models for dysphagia risk identification for adult patients who required endotracheal intubation within an ICU.

Methods

In this cohort study, two large healthcare databases were linked using deterministic logic. All participants received invasive mechanical ventilation in an ICU. Several machine learning model candidates were explored. Insights into the model decision-making have been provided using SHapley Additive exPlanation values.

Results

A total of 59 811 patients from 42 sites were included in the study. The top five most influential factors in determining the presence or absence of dysphagia at a cohort level were duration of mechanical ventilation, age, cardiac admission, neurological admission, and Acute Physiology and Chronic Health Evaluation III score.

Conclusion

There is a promising prospect of machine learning in dynamic dysphagia risk screening, which we propose should be considered for clinical use in the future. The patient-specific influence of each risk factor in determining the presence or absence of dysphagia highlights the importance of determining risk based on the individual patient's unique combination of risk factors, and not on cohort means, as has been done previously.