.png){kind=icon}
.png){kind=icon}
Return
AI-Enhanced Early Detection of Abnormal Tibial Fracture Healing Trajectories Using Temporal Deep Learning: A Physics-Aware RF Sensing Approach
.png){kind=icon}
OA
AI
A
A
E
P
N
I
V
R
DOI：10.1109/jsen.2026.3706261.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Noninvasive longitudinal monitoring of fracture healing remains clinically important, but current radiographic follow-up is episodic, ionizing, and often insensitive to early biological change. This article reports an initial simulation-based proof of concept for classifying tibial fracture-healing trajectories from radio frequency (RF) impedance measurements at 2.4 GHz using temporal deep learning. No phantom, experimental, or clinical data are used; all performance estimates should, therefore, be interpreted as computational feasibility bounds within the stated 2-D-finite-difference time-domain (FDTD) modeling assumptions and not as estimates of clinical performance. A bidirectional long short-term memory (BiLSTM) classifier was trained on 9800 longitudinal samples generated from 460 virtual patients with anatomical variability and measurement-noise augmentation. A central methodological result is that a three-class taxonomy (normal, nonunion, and abnormal) is more physically defensible than a conventional four-class taxonomy under the present sensing geometry. The four-class formulation failed to separate delayed union from malunion: delayed union recall was 0%, and no feature among 718 RF descriptors significantly discriminated delayed union from malunion. This supports grouping them as an abnormal trajectory that requires confirmatory imaging rather than claiming subtype diagnosis from RF data alone. Under patient-level fivefold cross-validation repeated across ten random seeds (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {n}={50}$ </tex-math></inline-formula> runs), the three-class BiLSTM achieved 81.9% accuracy (95% CI: 80.8%–82.9%) and 79.4% macro-<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$F1$ </tex-math></inline-formula>-score. Within the simulation-only retrospective truncation experiment, abnormal-class recall was 100% from Week 2 (Clopper–Pearson 95% CI: 90.3%–100.0% per fold), corresponding to zero simulated abnormal false negatives under the training and testing distribution. This result is not presented as clinically validated sensitivity: prospective deployment would need causal or variable-length models and validation against hardware nonidealities, sensor repositioning, motion, tissue heterogeneity, cast effects, calibration drift, manufacturing tolerances, anatomical complexity, and full 3-D anatomical variability. The BiLSTM also outperformed LSTM (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$+{1}.{9}\%$ </tex-math></inline-formula>, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {p}={0}.{012}$ </tex-math></inline-formula>) and GRU (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$+{5}.{2}\%$ </tex-math></inline-formula>, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {p}\lt {0}.{0001}$ </tex-math></inline-formula>) baselines while retaining millisecond-scale inference. The study, therefore, establishes a physics-aware, simulation-bounded framework for RF-based fracture-healing triage and motivates subsequent phantom, ex vivo, and prospective clinical validation.
Keywords:
Bidirectional long short-term memory (BiLSTM)
deep learning
early detection
finite-difference time-domain (FDTD) simulation
fracture healing
noninvasive monitoring
radio frequency (RF) sensing
sensitivity analysis
tissue dielectric variability
trajectory classification
Journal
IF:
4.5
Papers:
2.1W
Citations:
7.3W
Organization
Cited Papers