Abstract
Abstract
Background
Nursing assessments capture functional status and nursing-sensitive risks that are incompletely represented by diagnostic and administrative data. Integrating these dimensions may support multidimensional risk stratification. This study aimed to identify nursing assessment–derived risk profiles in adult inpatients, develop a VALENF Instrument–based classifier for automated profile assignment, and assess the profile-specific clinical equivalence of classifier-assigned profiles in an independent cohort.
Methods
This retrospective single-centre study included a development cohort of 2666 episodes and an independent cohort of 8218 episodes. Unsupervised clustering using age, sex, Charlson Comorbidity Index, and complete Barthel, Braden, and Downton assessments generated the reference profiles. Logistic regression, random forest, support vector machine, k-nearest neighbours, and ensemble models were evaluated across 1000 repeated 80/20 train–test splits. The final model was retrained on the complete development cohort and applied unchanged to the independent cohort. Profile distribution, ordinal gradients, and profile-specific clinical equivalence were assessed using Spearman correlations, bootstrap confidence intervals, and clinical margins.
Results
Five ordered profiles were identified (1: low-risk, 2: low-risk clinical-burden, 3: moderate-risk, 4: high-risk, 5: very-high-risk). Mean macro-F1 ranged from 0.819 to 0.842 across individual classifiers. Random forest was selected because it combined adequate performance (macro-F1 0.825; 82.43% correct assignments) with the lowest net imbalance towards lower-risk assignments. In the independent cohort, redistribution occurred between the two lower-risk profiles, whereas the remaining profiles had similar proportions. The ordinal gradient was preserved, and 25 of 30 profile-specific comparisons (83.3%) met the clinical-equivalence criterion.
Conclusions
Nursing assessment data supported five clinically interpretable, multidimensional risk profiles. A VALENF-based Random Forest classifier enabled automated assignment with adequate internal performance, while the independent cohort supported preservation of the aggregate clinical profile structure. Multicentre studies with complete reference assessments and prospective evaluation are required before clinical implementation.