Assessing Hospital Performance After Percutaneous Coronary Intervention Using Big Data.

Area of Science:

• Health Services Research
• Biostatistics
• Health Informatics

Background:

• Risk adjustment is crucial for comparing hospital outcomes but faces challenges with numerous confounders.
• Optimal strategies for risk adjustment are evolving, especially with the advent of big data.

Purpose of the Study:

• To compare conventional regression models with advanced methods for assessing hospital excess mortality risk.
• To evaluate the impact of confounder selection strategies (clinically informed vs. machine learning) on risk estimates.

Main Methods:

• Assessed 30-day excess mortality risk for percutaneous coronary intervention patients across 24 Massachusetts hospitals.
• Compared logistic regression with augmented inverse probability weighting and targeted maximum likelihood estimators.
• Utilized elastic net penalized regression for machine-learning-based confounder selection.

Main Results:

• Hospital excess risk estimates varied across methods and confounder sets, impacting outlier classification.
• Increasing confounders from 11 to 225 increased estimation uncertainty by an average of 62%.
• Agreement between methods ranged from fair (κ=0.39) to perfect (κ=1).

Conclusions:

• Modern causal inference techniques are essential for leveraging big data in hospital performance assessments.
• Minimizing bias through appropriate methods and confounder selection is key for accurate hospital outcome comparisons.