Deep learning approaches to predict femoral neck T-score and osteopenia/osteoporosis status from wrist accelerometry data.

Abstract

Osteoporosis is a prevalent condition with substantial health and economic implications. Although physical activity is associated with bone health, the specific temporal activity patterns related to bone mineral status remain insufficiently understood. This proof-of-concept study evaluated whether convolutional neural networks (CNNs) applied to Gramian Angular Field (GAF) representations of MIMS-derived wrist accelerometry data could predict femoral neck T-score and classify normal bone status versus osteopenia/osteoporosis. Data from 2504 participants in NHANES 2013-2014 were analyzed, including DXA-derived femoral neck measures, age, sex, body mass, and 7 days of wrist-worn accelerometry. Minute-level MIMS data were averaged into 10-min intervals and transformed into seven-channel GAF images. VGG-like and ResNet-like CNNs were trained from scratch using an internal validation set for early stopping and evaluated on a held-out test set. In classification, the VGG model achieved the best CNN performance, with test accuracy of 71.3%, sensitivity of 61.5%, specificity of 77.5%, F1-score of 0.625, and ROC-AUC of 0.780. In regression, the VGG model also performed best among CNNs, with test R2 of 0.305, MSE of 1.095, RMSE of 1.047, and MAE of 0.789. Demographic-only baseline models showed comparable or slightly better performance, indicating that age, sex, and body mass carried substantial predictive information. These findings support the feasibility of GAF-CNN modeling of accelerometry-derived activity patterns for bone-health research, while emphasizing the exploratory nature of the approach and the need for further validation.