Machine Learning and Deep Learning Approaches to Quantify Respiratory Distress Severity and Predict Critical Alarms

摘要

Respiratory distress (RD) is often the premonition and accompanying symptom of many critical conditions that may eventually lead to mortality among patients admitted in intensive care units (ICUs). The contemporary monitoring and alarm systems often fail to give timely and assertive alert of dangerously soaring RD, thus obscuring accurate prognosis. These systems do not capture the information provided by multiple parameters, which is important to predict and track patient health deterioration due to RD over time. Also, the repeated occurrence of false alarms in present systems leads to alarm fatigue. Our method addresses these issues by quantifying RD condition with a ‘Severity Index (SI)’ based on trend and value of respiratory rate (RR) and peripheral capillary oxygen saturation (SpO2) for 24 hours segment in a streaming monitor arrangement. We work on 912 records extracted from MIMIC-III Clinical and Waveform Database. We mapped our task as a classification problem and explored multiple machine learning and deep learning models in order to propose the best solution. The trend and value features were used for classification to train logistic regression, decision tree, support vector machine and multi-layer perceptron for quantification of RD severity on segment. We also use convolutional neural network (CNN) and long-short term memory (LSTM) for segment classification since they have capability to capture the temporal pattern of RD. All the models gave AUC (Area Under Curve for Receiver Operating Characteristic (ROC)) either close to or above 0.90 for classification task. We later used these models for raising clinical alarm for RD. We conclude that CNN model is marginally better than other models if we aggregate performance on all metrics. We can trigger an alarm based on SI of RD as evaluated by our CNN model instead of contemporary threshold based alarms of RR and SpO2. This RD alarm gives sensitivity and specificity of 86% and 85% respectively and achieves an average lead time of 5.5 hours without contributing to alarm fatigue. The dataset, code, trained models and the GUI are available at https://github.com/rohit-pardasani/RDQuantization.