Authors (including presenting author) :
CHAN NY (1,2), LEUNG E(3,4),LAW CB (1,2), WONG YF B (2), LEE A (4,6,8), GUAN JL (5), CHEN YHF(3), CHING CC (4), WONG CSM (4), LAM O(4), HE Y(4), YAU TYS (4), LIU Y(4), Hung CT (4), Yeoh EK (4), Hector TSANG H(6),CHOW YH (7)
Affiliation :
1. Department of Medicine & Geriatric, Princess Margaret Hospital, Kowloon West Cluster
2. Department of Medicine & Geriatric, North Lantau Hospital, Kowloon West Cluster
3. Department of Management Sciences, City University of Hong Kong
4. JC School of Public Health and Primary Care, The Chinese University of Hong Kong
5. EpitelligenceHK, Hong Kong
6. Department of Rehabilitation Science, Hong Kong Polytechnic University
7. Kwai Tsing Safe and Healthy City Association, Hong Kong
8. Centre for Health Education and Health Promotion, JC School of Public Health and Primary Care, The Chinese University of Hong Kong
Introduction :
The practice of clinical medicine can be enhanced by integrating public health insights extracted from population-level EHRs. For example, profiling high risk patients in the population may inform the design of more targeted care pathway. In addition, the systematic evaluation of the impact of acute and post acute services on the discharged population could potentially bring precision to discharge and transition care planning. We contribute to this literature by offering population-level insight on the different profiles of patients and the patterns and impacts of acute and post acute care utilization they received.
Objectives :
In the following, hybrid machine learning methodology was performed to analyze the population EHRs of a municipality to identify the consistency in the heterogeneity of 28-day rehospitalization outcomes and the systematic variability of post acute and ambulatory care utilization among patients sharing similar case-mixing and acute care utilization parameters.
Methodology :
The typical clinical and acute care needs profiles of the inpatient population of a regional medical system were identified using unsupervised machine learning model (i.e. Partition Around Medoids), and factors responsible for the heterogeneity of 28-day rehospitalization outcomes among those with similar case mixing and utilization parameters within homogeneous segment partitioned around typical patients in populations admitted from 2014 to 2019 were revealed by supervised machine learning models (i.e. Unbiased Recursive Partitioning with Surrogate Splitting).
Result & Outcome :
The Digestive, Respiratory, Circulatory, Nephrology and Urology, and Musculoskeletal (MSK) systems of diseases and disorders were consistently the Major Clinical Categories (MCC) of the medoids of the 50 64 and the 65+ populations admitted to municipal medical system every year between 2014 and 2019. However, only the medoids whose MCC were MSK were also consistent across five years at the case mix levels for both the 50 64 (inflammatory and reactive arthropathies; IRA) and the 65+ (Sprain, Strain, and Tendon and Joint Disorders; SSTJD) populations. In addition, among 50 64 and 65+ patients who admitted between 2014 and 2019 with clinical case mixing and acute care utilization parameters similar to that of the medoid case mix of IRA and SSTJID (hereafter, the IRA and SSTJID segments), respectively, their 28 day rehospitalization outcomes were heterogeneous even though their case mixing and acute care utilization parameters were relatively homogeneous. To elucidate the underlying factors responsible for the heterogeneity of 28 day rehospitalization outcomes in the IRA (50 64) and SSTJID (65+) segments each year between 2014 to 2019, supervised machine learning algorithm were performed. Conditional inference has revealed that PAC utilization was consistently the greatest contributor to 28 day rehospitalization outcome, even after adjusting for the same set of case mixing and acute care utilization parameters that were also selected to segment to population. Also consistently observed among the inpatient populations admitted every year from 2014 to 2019 was that, while different types of PAC utilized were associated with various likelihood of 28 day rehospitalization, those who lacked algorithm selected PAC (hereafter, the NS group) showed the highest 28 day rehospitalization rates compared to the rest of their segments. Also observed in the inpatient population admitted each year from 2014 to 2019 is that the NS group also reported significantly greater likelihood of having two or count of chronic diseases. Discussion: Applying hybrid machine learning to the EHRs of the municipal inpatient population admitted each year between 2014 and 2015, the current study has identified from patients of homogeneous diagnostic and utilization profiles the underlying factors that are consistently responsible for the heterogeneity of their
