OBJECTIVES: Hospital-acquired pressure injuries (HAPIs) result in over 60,000 deaths in the U.S. at a societal cost of $26 billion. Early detection of those at greatest risk could improve outcomes. The Braden Scale is a HAPI risk-assessment tool, with scores collected in patient charts. Our first objective was to develop a machine learning algorithm to predict future Braden score trajectories from electronic health records (EHR). Second, we applied constrained optimization to the database to improve predictive efficiency.

METHODS: We used a retrospective cohort of hospitalized patient EHRs between 2011 and 2014 to calibrate the algorithm. These data contained HAPI diagnosis not present-on-admission, patient-specific Braden scores for each 12-hour clinical shift up to the 12^th shift, and age. We used multilevel ordered logistic regression to create a predictive algorithm of HAPI risk, and applied constrained optimization to identify the optimal number of scores needed to efficiently and accurately predict future risk while minimizing overuse.

RESULTS: Our predictive algorithm using mixed-effects ordered logistic regression that was best-fit with no fewer than five Braden scores in order to anticipate categorical changes in HAPI risk to a high degree of accuracy (log-likelihood=-109,945.29). The best fit regression model was: y_ij=(14.6969+u_i0)+0.0149*x_ij+0.0231*Age_i+e_ij. Based on assessment of the area under the curve (AUC), there was a statistically significant improvement in predictive validity of HAPI risk with 6 Braden scores (p<0.05).

CONCLUSIONS: A series of 5-6 Braden scores can be used to accurately predict a patient’s future HAPI risk. This machine learning algorithm can be used to conserve hospital resources by directing preventive practices to patients of highest need. Two points of future exploration are: (A) whether the time required to wait for a 6^th score comes at a concerning incremental budget impact; and (B) whether inter-rater reliability between Braden scores from separate shifts adds more parameter uncertainty than is worth the wait.