Machine learning using institution-specific multi-modal electronic health records improves mortality risk prediction for cardiac surgery patients.

Background: The Society of Thoracic Surgeons risk scores are widely used to assess risk of morbidity and mortality in specific cardiac surgeries but may not perform optimally in all patients. In a cohort of patients undergoing cardiac surgery, we developed a data-driven, institution-specific machine learning-based model inferred from multi-modal electronic health records and compared the performance with the Society of Thoracic Surgeons models. Methods: All adult patients undergoing cardiac surgery between 2011 and 2016 were included. Routine electronic health record administrative, demographic, clinical, hemodynamic, laboratory, pharmacological, and procedural data features were extracted. The outcome was postoperative mortality. The database was randomly split into training (development) and test (evaluation) cohorts. Models developed using 4 classification algorithms were compared using 6 evaluation metrics. The performance of the final model was compared with the Society of Thoracic Surgeons models for 7 index surgical procedures. Results: A total of 6392 patients were included and described by 4016 features. Overall mortality was 3.0% (n = 193). The XGBoost algorithm using only features with no missing data (336 features) yielded the best-performing predictor. When applied to the test set, the predictor performed well (F-measure = 0.775; precision = 0.756; recall = 0.795; accuracy = 0.986; area under the receiver operating characteristic curve = 0.978; area under the precision-recall curve = 0.804). eXtreme Gradient Boosting consistently demonstrated improved performance over the Society of Thoracic Surgeons models when evaluated on index procedures within the test set. Conclusions: Machine learning models using institution-specific multi-modal electronic health records may improve performance in predicting mortality for individual patients undergoing cardiac surgery compared with the standard-of-care, population-derived Society of Thoracic Surgeons models. Institution-specific models may provide insights complementary to population-derived risk predictions to aid patient-level decision making.

Comparison of total joint arthroplasty care patterns prior to the Covid-19 pandemic and after resumption of elective surgery during the Covid-19 Outbreak: A retrospective, large urban academic center study.

BACKGROUND: After the suspension of elective surgeries was lifted in June 2020 in New York State, challenges remained regarding coordination of total joint arthroplasty (TJA) cases. Using the experience from a high-volume health system in New York City, we aimed to describe patterns of care after resumption of elective TJA. METHODS: We retrospectively assessed 7,699 TJAs performed before and during the COVID-19 pandemic. Perioperative characteristics and clinical outcomes were compared between TJAs based on time period of performance: 1) pre-pandemic (PP, June 8th-December 8th, 2019), 2) initial period post-resumption of elective surgeries (IR, June 8th-September 8th, 2020), and 3) later period post-resumption (LR, September 9th-December 8th, 2020). RESULTS: LOS > 2 days (83%, 67%, 70% for PP, IR, LR periods respectively) and discharge rates to post-acute care (PAC) facilities were lower during the pandemic periods (ORIR vs. PP: 0.48, 95% CI: 0.40-0.59, p < 0.001; ORLR vs. PP: 0.63, 95% CI: 0.53-0.75, p < 0.001). Compared to the pre-pandemic period, the risk for 30-day readmission was lower during the IR period (OR: 0.62, 95% CI: 0.40-0.98, p = 0.041) and similar during the LR period (OR: 0.96, 95% CI: 0.65-1.41, p = 0.832). CONCLUSIONS: Despite decreased LOS and discharge to PAC for TJAs performed during the pandemic, 30-day readmissions did not increase. Given the increased costs and lack of superior functional outcomes associated with discharge to PAC, these findings suggest that discharge to PAC facilities need not return to pre-pandemic levels.

Adherence to fluid resuscitation guidelines and outcomes in patients with septic shock: Reassessing the "one-size-fits-all" approach.

OBJECTIVE: The Surviving Sepsis Campaign and Centers for Medicare and Medicaid Services (CMS) Severe Sepsis and Septic Shock Management Bundle (SEP-1) recommend rapid crystalloid infusion (≥30 mL/kg) for patients with sepsis-induced hypoperfusion or septic shock. We aimed to assess compliance with this recommendation, factors associated with non-compliance, and how compliance relates to mortality. DESIGN: Retrospective, observational study. SETTING: 1136-bed academic and 235-bed community hospital (January 2015-June 2016). PATIENTS: Patients with septic shock. INTERVENTIONS: Crystalloid infusion (≥30 mL/kg) within 6 h of identification of septic shock as required by CMS. MEASUREMENTS: Associations with compliance and how compliance associates with mortality; odds ratios (OR) and 95% confidence intervals (CI) reported. MAIN RESULTS: Overall, 1027 septic shock patients were included. Of these, 486 (47.3%) met the 6-hour 30 ml/kg fluid requirement. Compliance was lower in patients with congestive heart failure (CHF) (40.9%), chronic kidney disease (CKD) (42.3%) or chronic liver disease (38.5%) and among those that were identified in the inpatient setting (35.4%) rather than in the emergency department (51.7%). When adjusting for relevant covariates, compliance (compared to non-compliance) was not associated with in-hospital mortality: OR 1.03 CI 0.76-1.41. CONCLUSIONS: These findings question a "one-size-fits-all" approach to fluid administration and performance measures for patients with sepsis.

The Impact of Hospital Size on CMS Hospital Profiling.

BACKGROUND: The Centers for Medicare & Medicaid Services (CMS) profile hospitals using a set of 30-day risk-standardized mortality and readmission rates as a basis for public reporting. These measures are affected by hospital patient volume, raising concerns about uniformity of standards applied to providers with different volumes. OBJECTIVES: To quantitatively determine whether CMS uniformly profile hospitals that have equal performance levels but different volumes. RESEARCH DESIGN: Retrospective analysis of patient-level and hospital-level data using hierarchical logistic regression models with hospital random effects. Simulation of samples including a subset of hospitals with different volumes but equal poor performance (hospital effects=+3 SD in random-effect logistic model). SUBJECTS: A total of 1,085,568 Medicare fee-for-service patients undergoing 1,494,993 heart failure admissions in 4930 hospitals between July 1, 2005 and June 30, 2008. MEASURES: CMS methodology was used to determine the rank and proportion (by volume) of hospitals reported to perform "Worse than US National Rate." RESULTS: Percent of hospitals performing "Worse than US National Rate" was ∼40 times higher in the largest (fifth quintile by volume) compared with the smallest hospitals (first quintile). A similar gradient was seen in a cohort of 100 hospitals with simulated equal poor performance (0%, 0%, 5%, 20%, and 85% in quintiles 1 to 5) effectively leaving 78% of poor performers undetected. CONCLUSIONS: Our results illustrate the disparity of impact that the current CMS method of hospital profiling has on hospitals with higher volumes, translating into lower thresholds for detection and reporting of poor performance.

Unplanned Readmission to the Pediatric Cardiac Intensive Care Unit: Prevalence, Outcomes, and Risk Factors.

BACKGROUND: Factors leading to cardiac intensive care unit (CICU) readmission and the impact on mortality have yet to be well delineated. We sought to define the prevalence and outcome for unscheduled CICU readmission. Secondary objectives were to identify indications and risk factors for unscheduled CICU readmission. METHODS: Retrospective analysis of prospectively collected registry data at a tertiary care children's hospital. Pediatric and adult patients with congenital and acquired heart disease who survived to initial CICU discharge were included. Patients with unexpected return to the CICU for acute change in clinical status were defined as unscheduled readmissions. RESULTS: Of the 645 discharges that met inclusion criteria, 37 resulted in unplanned readmission to the CICU. Patients requiring unscheduled readmission had higher mortality rates (16.2% vs 0.5%, P < .0001). Cardiac symptoms were the most common reason for readmission. On multivariate analysis, genetic anomaly (P = .001) and longer length of stay (LOS) during the index CICU admission (P = .01) were independently associated with readmission. For surgical patients, genetic anomaly (P = .001), single-ventricle anatomy (P = .05), and longer surgical support time (P < .001) were independently associated with readmission. CONCLUSION: Unscheduled readmission to the CICU within the same hospitalization was uncommon but associated with a higher mortality rate. Genetic anomaly and longer initial LOS were important risk factors for the entire cohort. Single-ventricle anatomy and longer intraoperative course were risk factors for surgical readmissions.