Scholarlink

Background <p>Patient Length of Stay (LOS) is a critical indicator of hospital efficiency, impacting resource allocation, costs, and care quality. Predicting LOS is especially challenging in resource-limited settings due to data skewness, outlier sensitivity, and imbalanced datasets. Machine learning (ML) offers potential solutions but requires validation in low-resource contexts.</p> Objective <p>To develop and evaluate ML models for predicting prolonged LOS (≥ 6 days) at Debre Markos Comprehensive SpecializedHospital, Ethiopia, using electronic health records (EHRs) to optimize operational planning.</p> Methods and materials <p>A retrospective analysis of 27,268 patient records (2017–2024) from four hospital wards was conducted. Data preprocessing included feature engineering, one-hot encoding, and addressing class imbalance (20,768 “Not Prolonged” vs. 6,500 “Prolonged”) using the Synthetic Minority Oversampling Technique (SMOTE). Eight ML models (Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, MLP, SVM, KNN, Naïve Bayes) were trained (80% training, 20% testing with 5-fold cross-validation) and evaluated using accuracy, precision, recall, F1-score, and AUC-ROC. Feature importance was assessed via permutation analysis.</p> Results <p>Key demographics: 62.3% rural patients, peak age 29–49 years (26.8%), with a sharp hospitalization decline post-2021 (regional conflict impact). SMOTE significantly improved model performance, particularly in terms of recall and F1-scores. Gradient Boosting outperformed others (AUC: 0.70, Accuracy: 81%, Precision: 72%, Recall: 63%, F1-score: 0.76). Top predictors of prolonged LOS were ward type (Medical/Surgical/Pediatrics), age (29–65 years), pregnancy-related conditions, respiratory/digestive diagnoses, and rural residence.</p> Conclusion <p>Gradient Boosting with SMOTE effectively predicts prolonged LOS in resource-constrained settings using EHR data. Implementation of ML-driven LOS prediction can enhance bed management, staffing, and resource allocation, aligning with Ethiopia’s Health Sector Transformation Plan (HSTP) goals for efficient healthcare delivery. Future work should focus on real-time validation and integration into hospital systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning predicts prolonged patient length of stay in a resource constrained Ethiopian hospital

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月4日

查看全文
  • Abraham Keffale Mengistu,
  • Kerebih Getinet,
  • Tefera Alemayehu,
  • Kelemua Aschale Yeneakal,
  • Andualem Enyew Gedefaw,
  • Abraham Teym,
  • Bekalu Endalew,
  • Bayou Tilahun Assaye

摘要

Background

Patient Length of Stay (LOS) is a critical indicator of hospital efficiency, impacting resource allocation, costs, and care quality. Predicting LOS is especially challenging in resource-limited settings due to data skewness, outlier sensitivity, and imbalanced datasets. Machine learning (ML) offers potential solutions but requires validation in low-resource contexts.

Objective

To develop and evaluate ML models for predicting prolonged LOS (≥ 6 days) at Debre Markos Comprehensive SpecializedHospital, Ethiopia, using electronic health records (EHRs) to optimize operational planning.

Methods and materials

A retrospective analysis of 27,268 patient records (2017–2024) from four hospital wards was conducted. Data preprocessing included feature engineering, one-hot encoding, and addressing class imbalance (20,768 “Not Prolonged” vs. 6,500 “Prolonged”) using the Synthetic Minority Oversampling Technique (SMOTE). Eight ML models (Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, MLP, SVM, KNN, Naïve Bayes) were trained (80% training, 20% testing with 5-fold cross-validation) and evaluated using accuracy, precision, recall, F1-score, and AUC-ROC. Feature importance was assessed via permutation analysis.

Results

Key demographics: 62.3% rural patients, peak age 29–49 years (26.8%), with a sharp hospitalization decline post-2021 (regional conflict impact). SMOTE significantly improved model performance, particularly in terms of recall and F1-scores. Gradient Boosting outperformed others (AUC: 0.70, Accuracy: 81%, Precision: 72%, Recall: 63%, F1-score: 0.76). Top predictors of prolonged LOS were ward type (Medical/Surgical/Pediatrics), age (29–65 years), pregnancy-related conditions, respiratory/digestive diagnoses, and rural residence.

Conclusion

Gradient Boosting with SMOTE effectively predicts prolonged LOS in resource-constrained settings using EHR data. Implementation of ML-driven LOS prediction can enhance bed management, staffing, and resource allocation, aligning with Ethiopia’s Health Sector Transformation Plan (HSTP) goals for efficient healthcare delivery. Future work should focus on real-time validation and integration into hospital systems.