Improving Haemorrhage Prediction in General Ward Patients by Incorporating Patterns of Change Derived From Routine...

To evaluate whether patterns of change derived from routine ward measurements provide additional predictive value beyond the most recent clinical measurements for haemorrhage prediction in general ward patients.

A retrospective cohort study.

Among 16,197 adult patients admitted to general wards in three university-affiliated hospitals in South Korea, 1152 (7.1%) experienced haemorrhage, whereas 15,045 (92.9%) did not. Three machine-learning algorithms evaluated three feature sets based on the most recent clinical measurements, patterns of change or both. Models were trained and internally validated using stratified five-fold cross-validation. Predictive performance was compared across feature sets using the area under the receiver operating characteristic curve, sensitivity, specificity and F1-score. Shapley Additive Explanations were used to interpret the best-performing model.

Models using both recent measurements and patterns of change showed small but statistically supported improvements in the area under the receiver operating characteristic curve compared with models using recent measurements alone. Shapley Additive Explanations identified neutrophil count, platelet count, white blood cell count, urine output and vital signs as key predictors.

Patterns of change derived from routine ward measurements provided complementary predictive information for haemorrhage prediction in general ward patients beyond the most recent clinical measurements. Routinely collected ward data may therefore contain additional predictive information not captured by the most recent values alone.

Machine-learning models combining recent clinical measurements with patterns of change may support risk stratification in general ward patients; prospective evaluation is needed to determine their clinical utility.

The study examined whether patterns of change improve haemorrhage prediction. These findings suggest that routinely collected ward data may provide additional predictive information beyond the most recent clinical measurements. The potential value of this information for nursing surveillance and timely intervention warrants prospective evaluation.

Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis-artificial intelligence (TRIPOD + AI).