Effective Clustering of Nursing Homes Using Unsupervised Machine Learning Focusing on Dementia and Mental Illness

Huiwen Xu, Orna Intrator, Zhengqing Ye, Eva Culakova, Chenwei Wu, Haosong Rao, John Bowblis

Abstract:

Research Objective:
About 65% to 90% of nursing home (NH) residents have mental health (MH) conditions including dementia, depression, and serious mental illness (SMI). It is unknown whether NHs tend to cluster around combinations of proportion of residents signaling NH specialization in care of these conditions. Identifying such NH clusters is important for examining how these clusters are staffed and whether staffing levels provide quality care, largely meeting the needs of their targeted populations. This study aims to 1) implement unsupervised machine learning to identify clusters of mental health conditions in NHs; 2) examine NH characteristics related to the clusters; 3) investigate the association of NH staffing and quality in each cluster.

Study Design:

NH-level data from 2009-2017 were extracted from the Certification and Survey Provider Enhanced Reporting (CASPER) and LTCFocUS.org. Unsupervised machine learning algorithm (K-means with Euclidean Distance) clustered NHs based on percent residents with dementia, depression, and SMI. NH staffing levels for each cluster, measured in hours per resident day (HPRD) were extracted for registered nurses (RNs), RNs with administrative duties (RNADs), licensed practical nurses (LPNs), certified nurse aides (CNAs), and social service staffing. Panel fixed-effects regression with staffing-cluster interactions was used to determine if staffing was associated with total health deficiency points, and if the association was different by cluster.

Population Studied:

Freestanding, privately-owned NHs without aberrant staffing levels, resulting in 110,463 NH-year observations from 14,671 unique NHs over the study period.

Principal Findings:

Three clusters were identified: low dementia and mental illness (Low Cluster); high dementia and mental illness (High Cluster); and high dementia and depression, but low SMI (Mixed Cluster). In 2009, 24.8% of facilities were in the Low Cluster, 20.7% in the High Cluster, and 54.5% in the Mixed Cluster. By 2017, these respective groups made up 47.3%, 27.4%, and 25.3%. Changes in clusters followed national trends of prevalence of dementia (47.1% to 44.9%), depression (52.0% to 38.0%), and SMI (24.3% to 33.5%). Compared to Low Cluster NHs, High Cluster NHs had lower nurse staffing; while Mixed Cluster NHs had lower RNADs and LPNs staffing, but higher RNs and CNAs staffing. Regression results showed that higher nurse staffing (except LPNs) was associated with significantly lower deficiency points in all clusters. One HPRD more RNADs and RNs were more strongly associated with fewer deficiencies in High Cluster NHs than Low Cluster NHs (marginal effect = -10.3, -9.0 for RNADs; marginal effect = -9.4, -6.3 for RNs, respectively). Social service staff were associated with significantly lower deficiency points only in the High Cluster.

Conclusions:

Unsupervised machine learning is an effective method of detecting previously unknown patterns of resident case-mix and staffing in NHs. Findings indicate that NHs that served residents with more severe MH issues had lower staffing levels; however, the association of staffing and quality among these NHs was stronger; greater availability of social service staff was more important for NHs serving more MH residents.

Implications for Policy or Practice:

Improving quality of NH care commensurate with changing population characteristics requires identification of clusters of important resident characteristics and assuring staffing levels are appropriate to their needs.