https://atlas.rsna.org/schemas/2025-11/model.json

Patient-specific Hip Arthroplasty Dislocation Risk Calculator (Explainable Multimodal ML)

https://pubmed.ncbi.nlm.nih.gov/36523643/

1.0

Cody C. Wyles (email as provided): ude.oyam@ydoC.selyW

Supported in part by the Mayo Foundation Presidential Fund and NIH grants R01AR73147 and P30AR76312.

Institutional review board approved; HIPAA compliant; waiver of informed consent.

Hybrid deep learning image feature extractor (EfficientNet-B4 backbone feeding Swin-B vision transformer) combined with clinical variables; top 10 transformer imaging features selected via XGBoost; final multimodal survival XGBoost model (with survival embeddings) on 10 imaging + 21 clinical features.

Predicts individualized 5-year risk (hazard) of dislocation after primary total hip arthroplasty; supports preoperative planning and surgical decision-making to mitigate risk.

Preoperative planning; clinical decision support systems.

Decision support; provides risk estimates and explainable feature attributions; clinicians retain decision-making.

Adults undergoing primary total hip arthroplasty; intended to estimate dislocation risk within 5 years using a preoperative AP pelvic/hip radiograph plus demographics, comorbidities, and surgical characteristics; use in preoperative clinical environment.

Latest preoperative anteroposterior pelvic or hip radiograph plus 21 encoded clinical features (demographics, comorbidities, surgical characteristics).

Provide the latest preoperative AP pelvic/hip radiograph and required clinical variables; review generated patient-specific 5-year dislocation risk and the matrix of risk estimates across modifiable surgical options (surgical approach, acetabular liner type, femoral head size); examine SHAP-based explanations for contributing factors.

Single-institution retrospective study; no external validation; potential generalizability concerns despite images from >120 devices over 20 years; model does not account for intraoperative factors such as implant positioning or soft-tissue management; relatively low diversity in acetabular liner types among dislocation cases; rare, multifactorial outcome limits performance; prospective validation needed.

Use as an adjunct for preoperative risk stratification and to assess how modifiable surgical choices may alter dislocation risk.

Internal development with 10-fold cross-validation; holdout test set (10% of cohort); image classifier trained/validated on 22,724 images from 129 devices; no external validation reported; preprocessing and training details provided (frameworks, hyperparameters).

Training performed in PyTorch (v1.10.0) on four NVIDIA A100 GPUs; batch size 128; 100 epochs with learning rate scheduling and EMA; runtime/energy use not reported.