Design and implementation of predictive models and a decision support system for chronic obstructive pulmonary disease patients management based on the pulse oximetry protocol

Abstract

Chronic Obstructive Pulmonary Disease (COPD) is a chronic lung disease characterized by persistent respiratory symptoms and airflow limitation, often leading to exacerbations which can have severe consequences if not addressed promptly. The objective of COPD treatment is to manage the individual worsening symptoms and their causes to improve COPD health and vital signs. To manage COPD, patients are normally categorized into prognosis stages according to symptoms severity and frequency and airflow limitation severity. COPD is usually monitored for the effectiveness of the selection and dosage of treatment under such GOLD stages. However, the latter have limited value in predicting death from COPD or predicting COPD exacerbation by observing the change of worsening symptoms over time. Therefore, GOLD stages may hinder better and immediate COPD therapy and healthcare. GOLD 2020 recommended the design of models that can monitor worsening symptoms dynamics for predicting the status of an individual COPD patient over time. Various research provided models that used symptoms dynamics for COPD exacerbation prediction. However, most of them used various parameters whose range of values may make their models output widely varying by making difficult COPD interpretation, for example by relating values of predictors of COPD exacerbation to prompt the selection of the therapy. Although, some symptoms and vital signs can provide the accuracy of predictive models for COPD, they may not correlate for example, for a specific COPD interpretation and therapy prompting. During this research, we designed models that consider individual symptom patterns for early and effective management of COPD and exacerbation (risk) prediction by utilizing only data derived from the pulse oximetry protocol. The related data used lead to COPD clinical reliance, interpretability and therapy prompting. The main aim was to work with pulse oximetry data to provide the state-of-the-art solutions when the latter is used for COPD risk prediction and management, by achieving COPD clinical reliance, interpretability and therapy prompting. By leveraging retrospective datasets containing symptoms patterns from patients, we proposed three machine learning models to promote COPD patients' health analysis over Electronic Health Record (EHR) by enabling clinicians, caregivers, and patients to identify early

2 indicators of exacerbations, facilitating timely intervention and personalized management strategies. For classification of exacerbations, we applied five classification models on parameters derived from pulse oximetry warning signs about COPD exacerbation, for the clinical reliance about the sensitivity, specificity and the predictive accuracy of those warning signs. XGBoost classification model outperformed other trained models by achieving an accuracy of 91.04%. For the prediction of risk of exacerbation, a DSS was designed on pulse oximetry derived data and symptoms patterns to analyze the problem of oxygen that may arise with worsening symptoms and associated exacerbations states. The SHAP interpreter was integrated with a RF pretrained model, to provide more insights on predictors of exacerbation states which may include SPO2 or oxygen. The oxygen level can be interpreted and adjusted according to worsening symptoms (exacerbation states) as guided by the individual GOLD stage. RF attained an accuracy of 95.3 by outperforming other trained models and was then integrated with a SHAP interpreter tool for the DSS design and COPD interpretability. Finally, the DSS was validated on 56 patients by attaining a concordance rate of 94.9%during a prospective study, when comparing the DSS with medically made decisions about exacerbation events. The third model was developed finally, to cope with frequent COPD exacerbations that may present during the disease course time. A LSTM model was applied on COPD patients’ data time series that contained symptoms patterns. One day data sampling rate was used to determine at which short-time interval, the early monitoring system for the therapy prompting about COPD subsequent exacerbations, could be based. The model achieved an accuracy of 85%, which can enhance the manual monitoring system that is utilized in pulse oximetry clinical practice, especially with hospitalized patients, when monitoring and diagnosing their exacerbations states and problem of low oxygen, as to avoid their causal to the subsequent exacerbations.