Abstract Body (Do not enter title and authors here): Background: Cardiogenic shock *CS) confers a high mortality (30-50%). A key limitation of existing CS mortality models is the reliance on static parameters measured at admission which ignore the rapidly evolving natural history of CS.
Hypothesis: A machine learning (ML) model that incorporates time series electronic health record (EHR) data including echocardiogram reports and invasive hemodynamics will accurately predict mortality in patients with cardiogenic shock
Methods: Patients with an ICD-10 code of cardiogenic shock were identified using electronic health data (EHR) from 2015 - 2019. The primary outcome was a composite of in hospital mortality or discharge to hospice. We incorporated EHR data obtained within the first 24 hours of admission including demographics, comorbidities, vital signs, lab values, medication administration, use of mechanical circulatory support (MCS), echocardiogram data, electrocardiogram reports, and hemodynamics from pulmonary artery catheter or right heart catheterization. For dynamic values such as lab values we incorporated six hour, 12 hour, and 24 hours trends as well as maximum and minimum values. We used a LightGBM model with 70/10/20 split for training, hyperparameter fine-tuning, and testing.
Results: We identified 5,048 total admissions. Median age was 65 years (IQR 52, 72) and 64% were male. Twenty five percent presented with acute myocardial infarction and 12% with cardiac arrest. The MCS distribution was: 17% intra-aortic balloon pump and 6% VA-ECMO. All-cause mortality was 32%. Our model demonstrated robust discrimination (AUROC 0.804), precision (0.699), and calibration (expected calibration error 3.2%). The five highest importance features were age at admission, mean serum albumin, mean glascow coma score, mean platelet value in the first six hours of admission, and total urine output. Our model performed favorably to a previously developed cardiogenic shock mortality risk scores (IABP SHOCK II: AUROC 0.71, AUPRC 0.55 in our cohort).
Conclusion: This ML CS mortality model incorporating dynamic, time series EHR data capturing the trajectory of CS in the first 24 hours after amission has high precision and discrimination. This ML model could be readily embedded into EHR systems to allow for rapid risk statification at the bed-side.