Radiomic Signatures From Conventional
MRI
Predict Recovery in Obstetric Brachial Plexus Palsy: A Machine Learning Approach in Infants
Gokcen Coban Cifci, Hanife Avcı, Ekim Gümeler, Meltem Karatas, Akın Üzümcügil, Tuzun Firat, Kader Karli Oguz, Gulgun Sengul ABSTRACT
Purpose
Obstetric brachial plexus palsy (OBPP) causes variable neurological outcomes, but early prognostic biomarkers remain limited. This study evaluated whether radiomic features extracted from routine T2‐STIR MRI can predict functional recovery in infants with OBPP, and whether combining these features with routinely available clinical predictors improves prognostic performance.
Methods
In this prospective single‐center study, 39 infants with OBPP (mean age 12.5 ± 5.7 weeks) underwent 3 T MRI during physiological sleep. Postganglionic lesions were semi‐automatically segmented on coronal T2‐STIR images, and 111 radiomic features were extracted per patient. After a minimum two‐year clinical follow‐up (mean 25.5 ± 2.0 months; range 24–32 months), patients were classified as having complete ( n = 17) or partial recovery ( n = 22). Feature selection was performed using the Boruta algorithm. Five supervised machine learning models were trained and internally optimized on the training set ( n = 28) using leave‐one‐out cross‐validation (LOOCV), and their performance was subsequently evaluated on an independent holdout test set ( n = 11). A clinical model (Narakas classification, birth weight, pseudomeningocele) and a combined clinical–radiomic model, summarized as a nomogram, were additionally developed.
Results
Eight radiomic features, including shape, first‐order, and texture metrics, differentiated recovery groups. Among the radiomics‐only models, SVM‐RBF performed best (training AUC = 0.929, F1‐score = 0.889; test AUC = 0.867, F1‐score = 0.750), followed by random forest (training AUC = 0.952; test AUC = 0.783); test‐set specificity was low for several classifiers. The combined clinical–radiomic model outperformed the radiomics‐only model (training AUC = 1.000; test AUC = 0.933, accuracy = 0.818), and the bias‐corrected calibration curve of the nomogram showed good agreement with observed outcomes (mean absolute error = 0.07).
Conclusion
Radiomic analysis of routine T2‐STIR MRI shows promise for predicting recovery in OBPP, and its combination with simple clinical predictors further improves discrimination. Given the small sample size and the performance drop between training and test sets, these findings should be regarded as preliminary and hypothesis‐generating, and require validation in larger, multicenter cohorts before clinical use.