A machine learning model has superior ability to predict 30-day mortality in patients with community-acquired pneumonia (CAP) compared with existing scoring systems, according to study findings published in CHEST.
Researchers conducted a derivation-validation retrospective study using a causal probabilistic network (CPN) to predict the risk of 30-day mortality in patients with CAP. The study comprised 4531 patients (derivation cohort) with CAP admitted to the Hospital Clinic of Barcelona in Spain between January 2003 and December 2016 and 1034 patients (validation cohort) admitted to the University Hospital la Fe of Valencia in Spain between January 2012 and December 2018.
The researchers modified the SepsisFinder (SeF) CPN that originally predicted mortality in patients with sepsis to predict mortality in patients with CAP via machine learning (SeF-ML). They validated the SeF-ML model via comparisons with 4 other clinical CAP scoring systems, including the Pneumonia Severity Index (PSI), the Sequential Organ Failure Assessment (SOFA), the quick Sequential Organ Failure Assessment (qSOFA), and the CURB-65 (confusion; urea, >7 mmol/L; respiratory rate ≥30/min; systolic blood pressure, <90 mm Hg and/or diastolic blood pressure, ≤60 mm Hg; age, ≥65 years) criteria.
The researchers calculated the area under the curve (AUC) to determine the ability of each scoring system to predict 30-day mortality in patients with CAP.
For patients in the derivation cohort (median age, 73 years; men, 60%), the SeF-ML model demonstrated the highest AUC (0.801), followed by the PSI (0.799), the CURB-65 (0.759), the SOFA (0.671), and the qSOFA (0.642).
Similar results were noted among patients the validation cohort (median age, 72 years; men, 62%). The SeF-ML model demonstrated an AUC of 0.826, indicating that its performance was not significantly different between the 2 cohorts (P =.051). In addition, the AUC of the SeF-ML model also was significantly higher than the qSOFA (AUC, 0.729; P =.005) and the CURB-65 (0.764; P =.03), and insignificantly higher than the SOFA (AUC, 0.771; P =.14). In contrast to its performance among patients in the derivation cohort, the AUC of the PSI was higher compared with the SeF-ML model among patients in the validation cohort (AUC, 0.830 vs 0.826), though the AUC was not significantly different between these 2 scoring systems (P =.92).
Study limitations include the lack of information on patients’ post-admission disposition, as well as differing baseline variables and clinical features of CAP between patients in the validation vs derivation cohorts.
According to the researchers, “[these] findings need further validation in other cohorts from different settings to assess the actual clinical utility of SeF-ML in predicting CAP prognosis.”
Cilloniz C, Ward L, Mogensen ML, et al. Machine-learning model for mortality prediction in patients with community-acquired pneumonia: development and validation study. Chest. Published online July 15, 2022:S0012-3692(22)01243-0. doi:10.1016/j.chest.2022.07.005