Prediction of community-acquired pneumonia outcome in admissions using machine learning

Community-acquired pneumonia (CAP) is a respiratory condition associated to high mortality in populations over 16, particularly critical in old patients. The accurate prediction of outcome directly reduces mortality. Currently, CAP outcomes are assessed with clinical scores like CURB65, based on symptoms that are nonspecific to the disease. Recent literature has shown that machine learning (ML) has the potential to improve outcome prediction, but the sparse and incomplete nature of the data represent a challenge for the development of models that can be implemented clinically.

This study aimed to determine whether ML models can support outcome prediction in patients admitted to hospitals with CAP using routinely collected and time-dependent data from Leicester hospitals. This, by conducting a cluster analysis, modelling mortality, CAP related severity outcomes and predicting URB65 with the forecast of vital signs, implementing a methodology that explores the characteristics involved in the modelling process.

Data comprised 9390 admissions in the training set, and 7892 in the validation Set, for thirty-four clinical variables (fifteen time-dependent). Results of clustering analysis revealed strong behavioural clusters for time-dependent data. A CAP mortality model reported AUC of 0.78 using a GRU classifier. Results also showed improvement in models when balancing classes of the target variable in the training set, and using time-series. Models for severity outcomes showed that potential re-segmentation of outcomes might benefit predictions, in spite of poor accuracy in these models (0.53 CURB65, 0.30 EWS and 0.17 LoS). And importantly when predicting URB65 accuracy of 0.85 was obtained when modelled using GRU. This approach might represent an opportunity to anticipate critical outcomes.

These results suggest that ML models using time-series in them can have big impact in the prediction of CAP outcome, from many perspectives: complementing current approaches in hospital settings like CURB65, prediction of mortality or forecasting the severity of patients via vital signs that have shown correlation with CAP outcome.