From the 2023 HVPA National Conference
Changyu Yin BS (University of Florida College of Medicine), Yu Huang BS, Zheng Feng PhD, Ron Ison MS, Liliana Bell MHA, Rulman Pebe BS, Roy Williams MS, Ray Opoku MSc, Jiang Bian PhD, Eric Rosenberg MD, Patrick Tighe MD, MS
Hip fractures remain a leading cause of morbidity and mortality in older adults. Prior analyses of outcomes related to postoperative recovery largely focus on patient factors, yet often fail to include important features related to the inpatient clinical environment. A key feature of any clinical environment is the social network of clinicians, from a wide variety of disciplines, caring for a given patient. Prior efforts to examine the impact of individual clinicians on perioperative outcomes have been challenged by numerous factors related to context and statistical methodology, including the need to consider random effects modeling and the nesting of individual clinician practice within a broader practice environment. Here, we use graph models to analyze clinical team networks on the day of surgery for patients receiving hip hemiarthroplasty.
Our primary objective was to determine whether clinical team networks were associated with prolonged postoperative length of stay in this cohort. Secondary analyses compared clinical network structures in patients with a normal versus prolonged postoperative length of stay.
This project was approved by the University of Florida IRB. The cohort consisted of adults aged ≥ 50 years who received hip hemiarthroplasty at UF Health Shands Hospital between 2014 and 2022. The primary outcome of interest was a prolonged length of postoperative hospitalization, defined as greater than the 75<sup>th</sup> percentile (i.e., 7.10 days) [1, 2]. Features consisted of patient clinical factors (e.g., comorbidities, surgery process durations, demographics, procedure type, etc.) and clinical team network measures (e.g., centrality, average degree connectivity, clustering coefficient, authority, page rank, density, etc). The clinical team network models provider-provider interactions (nodes are individual providers, and edge weights represent the number of interactions—two providers took care of the same patient during a surgical encounter—between two providers. Prediction models were constructed using statistical, machine learning, and ensemble methods, i.e., logistic regression, multi-layer perceptron, and XGBoost across various combinations of feature categories. Feature importance was assessed with SHAP – a technique for interpreting prediction models by quantifying the contribution of each feature towards the final prediction. Additionally, network features were compared between the normal versus prolonged length of stay groups using two-sided t-tests and two-sided Kolmogorov-Smirnov tests.
The cohort included 654 patients, where 164 patients have a prolonged length of postoperative hospitalization stay. The clinical team networks for the prolonged and non-prolonged cohorts are displayed respectively in Figure 1a and Figure 1b. As shown in Table 1, XGBoost using both clinical and network features, achieved AUC 0.964, Recall 0.833, Precision 1.0, F1-score 0.909, and accuracy of 0.939. Six of the top 10 features are related to the clinical team network. The only patient factor, BMI at admission, in the top-10 list ranked 7<sup>th</sup> (Figure 2). Table 2 demonstrates consistent differences in network measure values between the two outcome groups.
The network of clinicians providing care to hip fracture patients on the day of surgery are strongly associated with the patient’s postoperative length of stay. Future studies should incorporate characteristics related to clinical teams alongside patient factors. Further work is necessary to characterize the composition of clinical teams and clinician team factors associated with improved outcomes.