Enhancing early mortality prediction for sepsis-associated acute respiratory distress syndrome patients via optimized machine learning algorithm: development and multiple databases' validation of the SAFE-Mo.

BACKGROUND

Acute respiratory distress syndrome (ARDS) is associated with high mortality, with sepsis accounts for 31-34% of cases. Given the global burden of sepsis (508 cases per 100,000 person-years) and its association with 20% of all global deaths, early mortality prediction in patients with sepsis-associated ARDS is critical. This study developed and validated the Sepsis-associated ARDS Fatality Evaluation Model (SAFE-Mo), a machine learning model designed to predict early mortality in sepsis-associated ARDS patients, enabling earlier identification of high-risk individuals.

METHODS

Data were extracted from the Medical Information Mart for Intensive Care IV (MIMIC-IV, v3.0), eICU Collaborative Research Database (eICU CRD, v2.0), and Northwest ICU (NWICU, v0.1.0) using Structured Query Language. SAFE-Mo was constructed using machine learning algorithm (svmRadialSigma) focusing on median survival days among deceased patients as the primary outcome. The model's performance was validated externally using the MIMIC-IV and eICU CRD database and compared against four commonly used clinical risk assessment models (acute physiology score III (APSIII), simplified acute physiology score II (SAPS II), sequential organ failure assessment (SOFA), charlson comorbidity index (CCI)). Additionally, NWICU was used to further validate SAFE-Mo's generalization. Discrimination, calibration, and clinical utility were evaluated using area under the curve (AUC), Decision Curve Analysis (DCA), and calibration curves.

RESULTS

SAFE-Mo demonstrated superior predictive capability of early mortality compared to traditional models. It showed the largest reasonable risk threshold probability range and highest net benefit. Calibration curves indicated a slight overestimation of mortality risk overall. With our simple SAFE-Mo web page, SAFE-Mo can assist clinicians in identifying high-risk patients early, like patients with unusually high levels of lactate in sepsis-associated ARDS, assessing prognosis, and facilitating risk-adjusted comparisons of center-specific outcomes. Practical advantages include guiding personalized treatment strategies, determining the need for aggressive interventions, and optimizing resource utilization.

CONCLUSION

This study utilized the MIMIC-IV, eICU CRD, and NWICU databases to construct and validate a machine learning model, SAFE-Mo, which predicts early mortality in patients with sepsis-associated ARDS and outperforms traditional prediction models across all metrics. SAFE-Mo can guide clinicians to focus on critical indicators such as lactate, urine output, anion gap, and others, enabling appropriate measures to improve clinical outcomes for high-risk patients.