Development and Internal Evaluation of Models for Predicting Prolonged Hospitalisation Among Patients With Sickle Cell Disease in a Public University Hospital in Ghana
Year: 2026
Type: preprint
Study design: Observational Study
Abstract: Background: Prolonged hospitalisation is an important burden among patients with sickle cell disease (SCD), with implications for healthcare utilisation and patient and caregiver burden. Early prediction may support timely clinical and resource-management interventions. Objective: To identify factors associated with prolonged hospitalisation among patients with SCD and compare the predictive performance of multivariable logistic regression with five supervised machine-learning algorithms. Methods: This retrospective study included SCD inpatients aged ≥16 years admitted to the University of Ghana Hospital between January 2019 and December 2024. Logistic regression, random forest, XGBoost, support vector machine, and k-nearest neighbours models were developed using five-fold stratified cross-validation. Class weighting and SMOTE were applied within training folds. Model performance was evaluated using AUC-ROC, AUC-PR, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, and balanced accuracy. Results: Among 872 patients, 137 (15.7%) experienced prolonged hospitalisation. Acute chest syndrome (aOR=4.28; 95% CI: 2.31–7.94), malaria (aOR=3.77; 95% CI: 2.11–6.74), pneumonia (aOR=3.45; 95% CI: 2.02–5.90), and anaemia (aOR=2.71; 95% CI: 1.13–6.51) were independently associated with prolonged hospitalisation, whereas higher haemoglobin was protective (aOR=0.86; 95% CI: 0.78–0.96). Logistic regression achieved the highest AUC-ROC (0.692), F1 score (0.418) and balanced accuracy (0.677), while random forest achieved the highest AUC-PR (0.375). KNN had the highest sensitivity (59.3%) and XGBoost the highest specificity (87.2%). AUC-ROC confidence intervals overlapped substantially across models. Conclusion: Infection-related complications and anaemia were important predictors of prolonged hospitalisation, while higher haemoglobin was protective. ML models showed modest performance without clear superiority over logistic regression. Larger multicentre studies with external validation are warranted.
Source: medRxiv
Institutions: University of Ghana, The Salvation Army
Language: English
Subfield: Genetics (clinical)
Field: Medicine
Domain: Health Sciences
Keywords: sickle cell disease, prolonged hospital stay, hospital length of stay prediction, acute chest syndrome, malaria, pneumonia, anemia, hemoglobin, machine learning prediction models, random forest, XGBoost, logistic regression, risk factors, health care utilization, Ghana, SMOTE, class imbalance, infectious complications, caregiver burden, patient burden
Open Access status: green
OpenAlex ID: W7226028922
