Prediction of 1-year survival among elderly patients hospitalized in ICU for acute respiratory infection - 20/03/24

Intensive care unit (ICU) hospitalizations of elderly patients with acute respiratory infection have increased; however, we observed an important mortality during the ICU stay and the following year of discharge, giving doubts whether elderly patients benefit from ICU admission. We hypothesized that data-driven algorithms could be fueled by healthcare databases to build long-term prediction. Our objective was to assess the ability of machine learning (ML) algorithms to predict the one-year survival rate of elderly patients with severe respiratory infections. This assessment involved analyzing all relevant information from the French national hospital discharge database that was accessible prior to the admission.

A national 2013–17 cohort of patients≥80 y. o. with respiratory infection hospitalized in ICU was carried out, using hospital discharge databases. Patient characteristics and care healthcare trajectories before ICU were assessed within a 3-month period before ICU admission. Logistic regression, Random forest and XGBoosting are machine learning (ML) methods used to train predictive model. The accuracy, precision, the Matthews coefficient correlation (MCC) and the area under the curve (AUC) were measured to assess the ML models. A calibration plot was produced to measure the reliability of the model. Feature Importance (FI) were analyzed to interpret in a global way which covariates have an impact on the prediction and to explain the individual prediction. The LIME algorithm was constructed to identify the contribution of the 10 most influential features on prediction.

In total, 40,327 elderly patients were hospitalized in ICU for respiratory infection over the period. In total, 24,270 had a known vital status at 1year; 14,244 (58.7%) died during ICU stay or within the first year of discharge. Patients’ characteristics were: age 84 [82–87] y.o., male 19,708 (55.2%), SAPS II 33.7 [33.5; 33.9], invasive mechanical ventilation 9241 (38.1%). Logistic regression was the best ML model and discriminated the patient's outcome at one year with an AUC of 0.71 based on data collected within 3 months before the acute event leading to ICU. Example of individual prevision can be proposed thanks to the LIME algorithm.

We demonstrated that a ML model has the capability to forecast long-term outcomes for elderly patients with acute respiratory infections before their admission to the ICU. We believe that ML prediction will eventually enhance existing human expertise at the condition that physician are able to understand and challenge the features that drive the prediction. Indeed, we proposed, as a proof-of-concept, the use of a program that plays the role of the “explainer” of the ML model.