Analisi dei ricoveri causativi di reingresso con un algoritmo IA/machine learning: caratterizzazione clinica ed economica

[Machine translation] Early discharge represents a risk both for the patient, who finds himself in a setting inadequate to his current health needs, and for the healthcare company, which may incur a second hospitalization after a short time or a complaint that increasingly leads to a malpractice claim, with significant management costs. In fact, many regional health systems implement ad hoc controls on repeated admissions, which often result in economic losses. It is also known from the literature that patients on re-admission very frequently have worse clinical conditions than when admitted at the first admission, absorbing many more resources and resulting in a heavier DRG. Identifying these cases accurately and before discharge would make it possible to avoid unnecessary risks for patients and to optimize health resources, while avoiding a reduction in valuation., 15,093 admissions of patients over sixty-five under the ordinary regime of the Novara AOU for the year 2023 were analyzed in search of cases that followed a re-entry within 30 days for the same MDC. The disciplines of rehabilitation, psychiatry, oncohematology have been excluded. We have trained an extreme gradient boost (XGBoost) tree supervised artificial intelligence model to identify hospitalizations with potential “early discharge” and analyzed the importance of the features by calculating the shaps (ShaPley Additive Explanations) values, in order to obtain a ranking of the most relevant characteristics in causing potential re-entry. The dataset was split 80/20 and a 5 fold cross validation was conducted, after correcting the class imbalance. The analyses were conducted with R and RStudio and the packages: caret, xgboost and tidyverse., 301 re-entries have been identified with a total gross valuation (pre-reduction) of 1,818,202 euros and an average weight of 1.79. The final IA model for identifying causative hospitalizations reported an accuracy of 86.33% on the test set (95% CI: 84.36-88.14)). The most relevant characteristics according to the SHAP metric were the planned origin of the patient (SHAP average 0.159), the male gender (0.082), an age between 85 and 90 years (0.076) and the presence of the digestive system ultrasound diagnostic procedure (ICD procedure 8874, average SHAP value 0.068)., The IA model at the present time does not yet have sufficient performance for a practical implementation, but it only allows a better understanding of the characteristics of causative hospitalization. The research group is progressively implementing the main performance improvement techniques to allow predictive use.