Abstract
OBJECTIVE: Machine learning (ML) may allow for improved discernment of hemodynamics and oxygen delivery compared to standard invasive monitoring. We hypothesized that an ML algorithm could predict impaired delivery of oxygen (IDO(2)) with comparable discrimination to invasive mixed venous oxygen saturation (SvO(2)) measurement. METHODS: A total of 230 patients not on mechanical circulatory support (MCS) managed with a pulmonary artery catheter (PAC) were identified from 1012 patients admitted to a single cardiovascular intensive care unit (CVICU) between April 2021 and January 2022. Physiologic data were collected prospectively by the data analytics engine. Inadequate delivery of oxygen (IDO(2)) was defined as SvO(2) ≤50%. Fifty-four patients were used to train the model, which was then validated in 176 patients. Three simulated monitoring situations were constructed by downsampling the physiologic data set to exclude all SvO(2) sources (scenario A); all PAC data but allowing for SvO(2) values (scenario B); and all PAC data, including SvO(2) and cardiac index (CI) (scenario C). The ML platform then calculated the likelihood of IDO(2) for rolling 30-minute intervals and compared these values against the gold standard SvO(2) values using receiver operating characteristic (ROC) curve analysis to establish discriminatory power. RESULTS: A total of 1047 laboratory-validated SvO(2) values were collected for the validation group. The area under the ROC curve for the IDO(2) index was 0.89 (95% confidence interval, 0.87-0.91) with the full data set. When blinded to all PAC and SvO(2) sources, the AUC was 0.78 (95% confidence interval, 0.75-0.81). CONCLUSIONS: The IDO(2) index is capable of detecting SvO(2) ≤50% with good discriminatory function in non-MCS CVICU patients in a variety of monitoring situations. Further investigation of IDO(2) detection and clinical endpoints is needed.