Identifying 20 homogeneous clusters of acute patients discharged with nonspecific diagnoses through k-prototypes mixed data clustering

Le résumé fourni par la source

BACKGROUND: Patients discharged with nonspecific diagnoses after acute hospital care are frequent and represent potential diagnostic uncertainty at discharge. Adverse outcomes indicate missed diagnoses with a potential for improving patient safety. However, research and interventions are limited by population heterogeneity. We aimed to identify clusters of patients discharged with nonspecific diagnoses by employing unsupervised machine learning and to assess the risk of readmission and mortality of each cluster. METHODS: Observational, register-based study of emergency department arrivals discharged with nonspecific diagnoses (ICD-10: R and Z03 chapters) from March 2019 to February 2020 in Denmark. We applied partitional (k-prototypes) and hierarchical (agglomerative) clustering based on demographics, socioeconomics, comorbidities, administrative information, biochemistry, and 50 nonspecific discharge diagnosis groups. The risk of 30-day readmission and mortality after discharge was assessed as cumulative incidence for each cluster. RESULTS: We included 92,650 patients. A 20 clusters k-prototypes model best fitted our data. Clusters 1–5 were differentiated by no or limited biochemistry across different age and comorbidity patterns. Clusters 6–9 consisted mainly of young adults with low comorbidity, except Cluster 9 with notable neuropsychiatric and substance abuse comorbidities. Clusters 10–20 described the older patients: 10–14 with single comorbidities and 15–20 with substantial comorbidity of different cooccurring patterns. The risk of 30-day readmission and mortality ranged from 5% to 27% and 0% to 9% across clusters, respectively. CONCLUSION: Patients with nonspecific discharge diagnoses after acute hospital contacts can be grouped into 20 distinct clusters based on clinical, socioeconomic, administrative, and biochemical features. The clusters can be used to form delimited populations allowing for better and more individualized prediction models.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.