OBJECTIVES: Surgical procedures of low to medium complexity significantly benefit in quality, effectiveness, and efficiency of hospital resources from the implementation of outpatient surgery. Hospitals monitor their numbers of outpatient surgeries at different levels, but the comparison between hospitals is not straightforward. Outpatient surgery rates depend on a wide diversity of factors, such as hospital characteristics or inter-patient variation (including sociodemographic factors, pre-existing medical conditions, etc.) The objective was to develop a between hospital Adjusted Outpatient Surgery Index (AOSI) using machine learning (ML) methods with hospitalization episodes from more than 150 hospitals.

METHODS: The dataset comprised 1.6 million hospitalization episodes labelled as either outpatient or inpatient, extracted from the In-patient Database proprietary of IQVIA (APRGRD v36.0 and ICD-10 codes). ML techniques (gradient boosting algorithms) trained on hospital and patient variables (diagnosis, surgical procedures, comorbidities, socio-demographic information, etc.) were used to develop a model to estimate the probability of each medical episode being treated as outpatient. The AOSI methodology presented calculates AOSI scores as the ratio between observed and expected (predicted by the model) outpatient episodes at either hospital level, procedure category level, hospital department level, etc.

RESULTS: The modelling analysis confirmed that the information related to surgical procedures, age, comorbidities, and hospital type are key to understand the variability behind the decisions to treat a patient as outpatient. The ROC score measuring the performance of the model was 0.93 (ROC values range from 0 to 1), and the AOSI developed allowed inter-hospital evaluation taking both hospital and patients heterogeneity into consideration.

CONCLUSIONS: A new methodology supported by ML techniques was developed for the definition of the AOSI, allowing inter-hospital comparisons supported by high evaluation metrics associated to the model behind it.