Outcome Prediction in Burn Care: A Comparative Study of Machine Learning Algorithms on Structured Clinical Data
| Author | Abdolkhalegh Keshavarzi | en |
| Author | Ali Sam | en |
| Author | Nahid Abolpour | en |
| Author | Zahra Taghados | en |
| Author | Iliya Keshavarzi | en |
| Author | Fardin Ghanbari | en |
| Author | Marzieh Mansouri | en |
| Orcid | Abdolkhalegh Keshavarzi [0000-0001-6394-9019] | en |
| Orcid | Ali Sam [0009-0005-1109-8853] | en |
| Orcid | Nahid Abolpour [0009-0008-7649-9828] | en |
| Orcid | Zahra Taghados [0009-0002-0111-7334] | en |
| Orcid | Iliya Keshavarzi [0009-0008-3140-7926] | en |
| Orcid | Fardin Ghanbari [0009-0000-8222-7345] | en |
| Orcid | Marzieh Mansouri [0009-0006-3068-840X] | en |
| Accessioned Date | 2026-10-26T17:04:48Z | |
| Issued Date | 2026-07-31 | en |
| Abstract | Background: Early prediction of burn outcomes using structured clinical information available during prehospital care, triage, and early admission may support triage, counseling, and resource planning. Objectives: To compare five prespecified classifiers for the early prediction of final outcome, number of surgeries, intensive care unit (ICU) length of stay, and hospital length of stay using routinely collected prehospital and admission registry data. Methods: This retrospective single-center study used burn-registry records from Shiraz Burn Hospital for the period from 20 March 2020 to 20 March 2026. After records with final outcomes of transfer, discharge at the patient's own request, or unknown/not recorded were excluded, the analytic registry contained 4,127 records with 4,127 unique patient identifiers. Final outcome was modeled as recovery versus death. For the targets of number of surgeries and length of stay, deaths were excluded because death is a competing terminal event. Candidate predictors comprised demographic characteristics; referral and triage information; level of consciousness; vital signs and acute clinical findings; burn mechanism, intent, depth, total body surface area (TBSA), and anatomical regions; comorbidities; and prehospital treatment indicators. Target labels were not imputed. Logistic regression, decision tree, random forest, support vector machine (SVM), and k-nearest neighbors were compared using repeated stratified five-fold cross-validation, with macro-F1 as the primary model-selection metric, followed by evaluation on an independent 20% held-out test set. Results: The final outcome was recovery in 3,175 patients (76.9%) and death in 952 (23.1%). Among the 3,175 recovered patients included in the resource-use analyses, surgery categories were 0 surgeries in 48.1%, 1 in 14.0%, 2 in 10.3%, and > 2 in 27.6%; ICU-stay categories were no ICU/0 days in 74.9%, 1 - 7 days in 12.6%, and > 7 days in 12.5%; and hospital-stay categories were ≤ 7 days in 26.4%, 8 - 14 days in 33.2%, and > 14 days in 40.4%. In the development-set comparison, SVM achieved the highest macro-F1 for final outcome (0.895), whereas random forest performed best in predicting the number of surgeries, ICU length of stay, and hospital length of stay, with macro-F1 values of 0.466, 0.618, and 0.575, respectively. On the held-out test set, the selected SVM achieved a macro-F1 of 0.905 and a balanced accuracy of 0.932 for final outcome. For death prediction, sensitivity was 93.7% and positive predictive value was 79.2%. Random forest achieved held-out macro-F1 values of 0.459, 0.699, and 0.575 for number of surgeries, ICU length of stay, and hospital length of stay, respectively. Conclusions: Among the five evaluated classifiers, SVM provided the strongest performance for early prediction of final clinical outcome, whereas random forest performed best for the three resource-use outcomes. Early structured burn-registry variables enabled strong discrimination between recovery and death, including high sensitivity for mortality, whereas prediction of operative burden and hospital length of stay remained more challenging. These findings support further external and prospective validation before clinical implementation. | en |
| DOI | https://doi.org/10.5812/semj-173380 | en |
| URI | https://brieflands.com/journals/semj/articles/173380 | en |
| Keyword | Burn Injuries | en |
| Keyword | Machine Learning | en |
| Keyword | Patient Outcome Prediction | en |
| Keyword | Mortality | en |
| Keyword | Length Of Hospital Stay | en |
| Publisher | Brieflands | en |
| Title | Outcome Prediction in Burn Care: A Comparative Study of Machine Learning Algorithms on Structured Clinical Data | en |
| Type | Research Article | en |