Predicting surgical outcomes for chronic exertional compartment syndrome using a machine learning framework with embedded trust by interrogation strategies.
Journal:
Scientific reports
PMID:
34931008
Abstract
Chronic exertional compartment syndrome (CECS) is a condition occurring most frequently in the lower limbs and often requires corrective surgery to alleviate symptoms. Amongst military personnel, the success rates of this surgery can be as low as 20%, presenting a challenge in determining whether surgery is worthwhile. In this study, the data of 132 fasciotomies for CECS was analysed and using combinatorial feature selection methods, coupled with input from clinicians, identified a set of key clinical features contributing to the occupational outcomes of surgery. Features were utilised to develop a machine learning model for predicting return-to-work outcomes 12-months post-surgery. An AUC of 0.85 ± 0.08 was achieved using a linear-SVM, trained using 6 features (height, mean arterial pressure, pre-surgical score on the exercise-induced leg pain questionnaire, time from initial presentation to surgery, and whether a patient had received a prior surgery for CECS). To facilitate trust and transparency, interrogation strategies were used to identify reasons why certain patients were misclassified, using instance hardness measures. Model interrogation revealed that patient difficulty was associated with an overlap in the clinical characteristics of surgical outcomes, which was best handled by XGBoost and SVM-based models. The methodology was compiled into a machine learning framework, termed AITIA, which can be applied to other clinical problems. AITIA extends the typical machine learning pipeline, integrating the proposed interrogation strategy, allowing to user to reason and decide whether to trust the developed model based on the sensibility of its decision-making.
Authors
Keywords
Adult
Area Under Curve
Chronic Exertional Compartment Syndrome
Exercise
Fasciotomy
Female
Humans
Leg
Linear Models
Lower Extremity
Machine Learning
Male
Middle Aged
Military Personnel
Models, Statistical
Prognosis
Reproducibility of Results
Return to Work
ROC Curve
Sensitivity and Specificity
Support Vector Machine
Surveys and Questionnaires
Time Factors
Treatment Outcome
Young Adult