A newly designed, gradient-boosting machine learning model to predict clinical deterioration in pneumonia (CDiP) outperformed the pneumonia severity index (PSI) in predicting risk of in-hospital disease progression. Research on the model’s development and validation was presented at the American College of Chest Physicians (CHEST) 2022 Annual Meeting, held October 16 to 19, in Nashville, Tennessee.
Researchers aimed to design and validate a machine learning tool, dubbed “Clinical Deterioration in Pneumonia” (CDiP), which predicts early risk of in-hospital disease progression in community-acquired pneumonia (CAP). Researchers also compared the newly developed tool’s performance to that of the Pneumonia Severity Index (PSI).
The investigators conducted a retrospective single-center study that included 4383 adult patients hospitalized with community-acquired pneumonia from January 2009 to December 2020. The study cohort was randomly divided into 2 groups, with 75% of participants in the training group (n=3288) and 25% in the validation group (n=1095). The researchers used electronic health records to collect demographic, physiologic, and laboratory data gathered within the first 6 hours of admission.
In developing the model, investigators focused on predicting deterioration from the time of admission or within 6 hours. They defined deterioration using a modified version of the World Health Organization clinical progression scale that included (1) supplemental oxygen; (2) noninvasive ventilation or use of high-flow oxygen devices; (3) invasive mechanical ventilation or extracorporeal membrane oxygenation; and (4) death. The researchers used a Gradient Boosting Machine model and internal 5-fold cross validation to develop and fine tune their predictive model.
A total of 25% of patients in both the training (832) and validation (277) groups experienced disease worsening in-hospital. In the training cohort, the new CDiP tool showed better discrimination than the PSI (area under the receiver operating characteristic curve [AUC], 0.83 vs 0.59, respectively). Likewise, in the validation group, the CDiP tool achieved an AUC of 0.95 (threshold accuracy rate 91%; 95% CI, 89%-92%; specificity=0.96; sensitivity=0.76), outperforming the PSI, which had an AUC of 0.53 (threshold accuracy rate 76%; 95% CI, 73%-78%; specificity=0.99; sensitivity=0.05).
Researchers concluded “The gradient boosted model CDiP was superior in predicting the risk of in-hospital disease progression compared to the PSI and demonstrated adequate discrimination and sensitivity in patients with community-acquired pneumonia.” They further noted that “Our new machine learning model (CDiP) can facilitate early identification of hospitalized patients at high risk for clinical deterioration from community-acquired pneumonia.”
Disclosure: 1 study author declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.
Odeyemi Y, Lal A, Barreto EF, et al. Machine learning prediction of in-hospital disease progression in community-acquired pneumonia: derivation and validation of clinical deterioration in pneumonia (CDiP). Presented at: CHEST 2022 Annual Meeting; October 16-19, 2022; Nashville, TN.