Integrating Machine Learning With Resilience Models To Assist Hospital Resilience improvement

Abstract: Healthcare systems have become increasingly fragile due to growing rates of burnout, depression, distress, and subsequent workforce shortages since the COVID-19 pandemic. Although researchers and institutions urgently called for hospitals and healthcare staff to build resilience to withstand, adapt, recover, rebound, or even grow from adversity, stress, or trauma, the ambiguous relationships between individual and organizational resilience are one of the major obstacles to the development of resilience in hospitals. The COVID-19 pandemic amplified the intersection of individual resilience and system resilience. It provided a good opportunity to discover the weak signals of cognitive behavior of healthcare staff that were ignored before the pandemic. The Patient Safety Culture Survey (PSCS) and the Employee Satisfaction Survey(ESS) are mandatory requirements for hospitals by the hospital accreditation. The PSCS has 46 questions in 8 dimensions. It aims to examine the dimensional strengths and weaknesses of hospitals’ patient safety culture. The ESS (39 questions, 7 dimensions) aims to understand how healthcare workers are satisfied with policy, management, team, job, etc. Both survey data include employees' feelings and attitudes about the work system and workload but have been used only for their designed purposes. Thus, the study aims to identify individual and collective weak signals from routine hospital survey data to proactively support employee retention programs and strengthen hospital resilience promotion activities.The study setting is a medical center. It collected 2020 – 2022 PSCS data including the emotional exhaustion questions adopted from the Maslach Burnout Inventory, and ESS data. First, we applied machine learning(ML) to determine which questions in the PSCS and ESS were associated strongly with workers’ resignation or retention. Next, based on the concepts of the multilevel organizational resilience model and the interplay model of individual and organizational resilience, we accordingly marked each determined question corresponding to individual or organizational resilience elements as a signal map for further improvement activities. Finally, by departments, we screened out the riskiest units to make suggestions for employee care and system resilience improvement. 3076 records of PSCS and 4521 in ESS were analyzed. The XGBoost method combined with 10-fold cross-validation was adopted in ML. By using SHapley Additive exPlanations(SHAP), the SHAP value of PSCS and ESS questions greater than 0.05 were selected as the most meaningful signals to resignation and retention accordingly. PSCS: questions S12 (education & training courses organized by our hospital) and S30, and ESS: questions Q14 (I like my job very much), Q16, Q27, and Q35 were identified for the retention signals. PSCS: question 27 (communication channels and methods of the hospital) and S30, and ESS: question 13 (has a culture that fosters colleagues to learn from others’ mistakes) and Q24, Q37 of ESS were identified for resignation signals. The determined questions were accordingly linked with the elements of individual resilience and organizational resilience by departments for future resilience improvement.The study provides a valuable framework that integrates with ML analysis to utilize the dynamic relationships between individual and organizational resilience and assist in future hospital resilience development.

Keywords: resilience engineering, workforce shortage, organizational resilience, machine learning

DOI: 10.54941/ahfe1005688

Cite this paper:

​