AI-Assisted Failure Prediction on Postoperative X-Rays: A Head-to-Head Comparison with Consultant Trauma Surgeons

Abstract Background & Objectives: Surgical hardware failure after orthopaedic fracture fixation causes morbidity and rising cost. Immediate postoperative radiographs may carry subtle markers of mechanical failure. This study compared a convolutional neural network-based artificial intelligence model with three blinded trauma surgeons in predicting 12-month fixation failure and assessed the model's potential as a clinical decision-support adjunct. Material & Methods This prospective multicentre study analysed 466 patients (5–91 years; 59% male; 30% open fractures) with operatively treated humerus, femur, ulna, radius or tibia fractures and standardised immediate postoperative radiographs. Failure was a composite endpoint of hardware breakage, screw or plate migration, non-union, or malunion requiring revision within 12 months, adjudicated independently. A ResNet-50 CNN, fine-tuned on an independent 200-patient development cohort with no overlap with the 466-patient application cohort, was compared with three surgeons using sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and class-wise F1 score, with paired testing via McNemar's test. Results Of 466 patients, 104 had fixation failure (22.3%). The AI model achieved sensitivity 96.15%, specificity 98.90%, PPV 96.15%, NPV 98.90%, F1 (failure class) 96.15%, and accuracy 98.28% (458 of 466 correct). The strongest surgeon achieved sensitivity 91.35%, specificity 98.62%, PPV 95.00%, F1 (failure class) 93.14%, and accuracy 97.00% (452 correct). The AI significantly outperformed Surgeons 1 and 2 (exact binomial p < 0.0001 and p = 0.0001). The comparison with Surgeon 3 did not reach statistical significance (exact binomial p = 0.0703). Conclusions The AI model outperformed every participating trauma surgeon across all metrics and significantly outperformed two of three surgeons; its advantage over the most experienced surgeon did not reach statistical significance in this cohort. These findings support a possible role as a clinical decision-support adjunct rather than a clinician replacement, pending external prospective validation, subgroup analysis by anatomical site and fixation technique, and resolution of ambiguity in the composite failure endpoint used to adjudicate the primary outcome.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11408305/v1
Primary Topic
Bone fractures and treatments
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

AI-Assisted Failure Prediction on Postoperative X-Rays: A Head-to-Head Comparison with Consultant Trauma Surgeons

Shah Dhaval, Gaurav Vatsa, Saurabh Suman
4 links
Research Square
Bone fractures and treatments
preprint

AI-Assisted Failure Prediction on Postoperative X-Rays: A Head-to-Head Comparison with Consultant Trauma Surgeons

Shah Dhaval, Gaurav Vatsa, Saurabh Suman
preprint en

Abstract

Abstract Background & Objectives: Surgical hardware failure after orthopaedic fracture fixation causes morbidity and rising cost. Immediate postoperative radiographs may carry subtle markers of mechanical failure. This study compared a convolutional neural network-based artificial intelligence model with three blinded trauma surgeons in predicting 12-month fixation failure and assessed the model's potential as a clinical decision-support adjunct. Material & Methods This prospective multicentre study analysed 466 patients (5–91 years; 59% male; 30% open fractures) with operatively treated humerus, femur, ulna, radius or tibia fractures and standardised immediate postoperative radiographs. Failure was a composite endpoint of hardware breakage, screw or plate migration, non-union, or malunion requiring revision within 12 months, adjudicated independently. A ResNet-50 CNN, fine-tuned on an independent 200-patient development cohort with no overlap with the 466-patient application cohort, was compared with three surgeons using sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and class-wise F1 score, with paired testing via McNemar's test. Results Of 466 patients, 104 had fixation failure (22.3%). The AI model achieved sensitivity 96.15%, specificity 98.90%, PPV 96.15%, NPV 98.90%, F1 (failure class) 96.15%, and accuracy 98.28% (458 of 466 correct). The strongest surgeon achieved sensitivity 91.35%, specificity 98.62%, PPV 95.00%, F1 (failure class) 93.14%, and accuracy 97.00% (452 correct). The AI significantly outperformed Surgeons 1 and 2 (exact binomial p < 0.0001 and p = 0.0001). The comparison with Surgeon 3 did not reach statistical significance (exact binomial p = 0.0703). Conclusions The AI model outperformed every participating trauma surgeon across all metrics and significantly outperformed two of three surgeons; its advantage over the most experienced surgeon did not reach statistical significance in this cohort. These findings support a possible role as a clinical decision-support adjunct rather than a clinician replacement, pending external prospective validation, subgroup analysis by anatomical site and fixation technique, and resolution of ambiguity in the composite failure endpoint used to adjudicate the primary outcome.

Research Square
All India Institute of Medical Sciences, Deoghar (IN)
Good health and well-being
Bone fractures and treatments
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon