This was a prospective, single-center cohort study of adults aged ≥18 years undergoing TKA by fellowship-trained arthroplasty surgeons at a high-volume orthopedic hospital in India between January 2025 and May 2025. Baseline demographic, comorbidity, and laboratory data were collected at the time of enrollment. Demographic variables were age and sex; procedural variables were surgical approach (conventional or robotic-assisted TKA) and operative laterality (right, left, or bilateral). Baseline laboratory values were hemoglobin, serum creatinine, sodium, and potassium. Documented comorbidities included hypertension, diabetes mellitus, thyroid disease, coronary artery disease, and chronic kidney disease, among others. These baseline characteristics are summarized in
Participants were recruited from patients scheduled to undergo TKA or who presented for TKA at the study site during the predefined enrollment period. Potential participants were assessed for eligibility according to the study inclusion and exclusion criteria. All eligible patients who expressed interest in participation were screened by the study staff, and participants were sequentially enrolled as they became available throughout the enrollment period. Sample size was determined by enrollment capacity during the defined study period rather than prospective power analysis.
The study was approved by the Institutional Ethics Committee (KIMS/IEC-RHR/2025/06/05) and conducted in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines for multivariable prediction models [
Paired infrared and EO data were acquired at postoperative visits for postoperative days 1, 15, 24, and 30, corresponding to routine clinic and physiotherapy appointments. In practice, acquisition timing varied with patient attendance, and all analyses used each scan’s recorded postoperative day rather than the nominal visit day. Imaging used a custom handheld device integrating an Intel RealSense D405 EO camera (640×480-pixel color) and a Seek Thermal Mosaic Core long-wave infrared radiometric sensor (7.8-14 µm; 240×320 pixels). Standardized acquisition was performed both during scheduled clinic visits and at patients’ homes during physiotherapist visits, capturing performance across clinic and in-home settings.
Thermal acquisition followed medical thermography guidance: emissivity ε≈0.98 with ambient reflection compensation, orthogonal framing, and consistent standoff distance, in accordance with IEC 80601-2-59 and ASTM E1933/E2847 [
Of the 1065 patients screened, 749 (70.3%) met the inclusion criteria; the remainder were excluded for incomplete medical records, data inconsistency or irregularities, inability to complete follow-up, or nonadherence to the study protocol. Thirty-four infection-positive patients were identified during follow-up, of whom 8 were excluded on predefined data-completeness grounds: one had a missing baseline serum potassium value, and 7 did not return for follow-up assessment, and confirmation of healing was not obtained through attendance or telecommunication attempts. This yielded 26 adjudicated infection-related complications in the analytic cohort (26/749, 3.5%). Of these, 25 out of 26 (96%) presented with wound dehiscence attributable to infection, of which 1 out of 26 (4%) additionally exhibited oozing blood with purulent discharge meeting the definition of a hematologic event. One patient (1/26, 4%) had confirmed PJI without visible dehiscence. The 749 patients contributed 2606 co-registered radiometric infrared and EO image pairs (84 complication and 2522 noncomplication pairs).
Development used radiometric data acquired at postoperative day 14 or later, comprising 409 patients (26 with infection-related complications and 383 controls) who contributed 1323 scans. This window was chosen because scan-level AUROC was 0.670 in the first postoperative week and exceeded 0.90 thereafter; within it, the number of scans available per knee was not itself predictive (area under the curve [AUC] 0.466), indicating that follow-up intensity did not act as a proxy for outcome. Separately, postoperative day was not used as a direct predictor, because complication-affected knees tended to be imaged later in the postoperative course, and day-of-scan would otherwise encode imaging schedule rather than pathology; thermal descriptors were instead expressed as deviations from expected uncomplicated healing at the same postoperative day, with the normative reference estimated from control scans in the training folds only. The classifier was trained and evaluated by 5-fold cross-validation with folds grouped by patient, so that both knees of a patient and all of that patient’s scans were assigned to a single fold. Every reported metric in this study was calculated out-of-fold; each patient was scored only by models fitted without that patient’s data, and results were averaged over 3 random seeds.
The classifier was framed as a combined detection and prediction task in which a knee was labeled positive if it developed a complication at any point during follow-up. Every scan of a complication-positive knee, including scans acquired before clinical onset, was assigned to the positive class, reflecting the intended clinical use in which any available scan should contribute to flagging an at-risk knee. Complementary learners were additionally trained under a stricter labeling scheme in which only scans acquired during a clinically present complication were positive; in this scheme, pre–onset scans were withheld from the training objective but were still scored at inference, so their observed values contributed to the patient-level aggregate and no score was imputed.
Each thermal scan was cropped to the region of interest (ROI) established by the EO pipeline, expanded by an 8% margin, and resized to 224×224 pixels so that imaging distance and framing did not influence the analysis. Each scan was encoded in 2 forms: as images, for convolutional analysis, and as a set of quantitative descriptors, for feature-based analysis. Both representations were derived from the radiometric thermal matrix within the ROI. Three complementary image encodings were used. The first was a 3-channel physics encoding comprising the thermal excess over ambient, the local SD over a 5×5 neighborhood, and the gradient magnitude. The second was a color map rendering of the thermal matrix, per-image minimum-maximum normalized. The third was a self-normalized encoding in which every channel was referenced to the knee’s own median and IQR, rendering it invariant to uniform shifts in scene temperature. The quantitative descriptors were physics-based thermal features organized into families: temperature and Stefan-Boltzmann radiant-exitance statistics, spatial gradient and local variability, the morphology and extent of localized hot regions, multiscale texture energy from a scattering transform across 4 spatial scales, band contrasts oriented along the incision axis, and each scan’s deviation from the patient’s own baseline, defined from the earliest complication-free scan and from expected uncomplicated healing at the same postoperative day.
Complementary learners were trained on these representations. Convolutional neural networks were trained on each of the 3 image encodings under each of the 2 labeling schemes; each used a classification head (a dropout layer followed by a linear layer) and was optimized with AdamW (learning rate 2×10-4 and a weight decay of 10-4) under a cosine schedule for 25 epochs at a batch size of 32, with class-balanced sampling to account for the low prevalence of complications and a cross-entropy objective. Augmentation was limited to horizontal reflection, translations of up to ±12 pixels, and scale changes between 0.92 and 1.08; vertical flips and rotations were not applied, as the images have a fixed anatomical orientation. Horizontal-flip test-time augmentation was used at inference. Two additional learners operated on the quantitative descriptors: a self-supervised image-embedding model reduced by principal component analysis and classified by logistic regression, and a logistic-regression model on the physics descriptors with postoperative-day normative scoring. The learners’ patient-level scores were combined by equal-weight rank averaging, a parameter-free step with no combination layer fitted to the labels, and each knee’s final score was calculated as the mean of its scores across scans.
Ambient temperature differed between the complication and control groups (27.4 °C versus 30.0 °C); because this difference was already present in scans acquired before clinical onset, it reflected the acquisition setting rather than the outcome. It was controlled in 3 ways: a self-normalized, ambient-invariant encoding that references every value to the knee’s own distribution so that a uniform shift in scene temperature cancels by construction, which as a standalone model reached an AUC of 0.950; measurement of each learner’s dependence on ambient temperature, which was low for the image-based learners; and regression of ambient temperature out of the patient-level score using control patients only, which yielded an AUC of 0.963, a conservative estimate of performance attributable to thermal pathology rather than to acquisition conditions.
Development used 2606 radiometric scans, partitioned into 2203 training scans and 403 validation scans. Thermal scans were labeled to create the binary segmentation masks of the operated knee as visible in the radiometric heatmap. The masks were used as ground truth. Training and validation were separated at the patient level so that no patient contributed images to both partitions, and the decision threshold was fixed on the training set before evaluation.
The anatomical landmark detector, the fine-tuned vision-language model (VLM), and the complication segmentation model were developed on cohort EO images expanded with data augmentation and integrated with publicly available datasets. Offline augmentation was performed using feature pasting, cross-feature mixing, background substitution and drift, and morphological operations was applied on the fly for training only and was excluded from all validation sets, which comprised unaltered images. Offline augmentation used no image generative AI (GenAI) tools and was done manually using Python (Python Software Foundation) and image-editing tools (
The landmark detector used 8635 training and 426 validation images, containing 914 annotated objects, at a fixed, untuned confidence threshold of 0.5. The dataset comprised 7 classes, including bandage, feet, hand, healed incision, incision, leg, and leg with bandage. A bounding box was created for each label as the ground truth. The validation set was held-out from training and screened against the full training set by perceptual hashing, with no validation image falling within a near-duplicate threshold of any training image. The VLM was fine-tuned on a multilabel corpus of EO wound images. Training annotation counts were 1379 infections, 1052 blood presence, 728 purulent discharge, 704 open wounds, 516 discharge or exudate, and 283 cellulitis instances, within a broader set of 17,712 images containing a surgical site. The held-out validation set contained 152 infection, 132 blood presence, 68 purulent discharge, 60 open wounds, 56 discharge or exudate, and 44 cellulitis-positive instances within 1956 surgical-site images. The validation set did not contain any augmented images. The complication segmentation model used 37,062 training and 2003 validation images, partitioned at the image level, and incorporated publicly available open-wound datasets. The model was trained on the segmentation of 4 classes, namely, open wound, granulation, infection-associated exudate (IAE), and necrosis. IAE is the superclass representing a single label for the segmentation of purulent discharge.
Illustrative offline data augmentation of an electro-optical wound image. Left, an original electro-optical frame from the study device showing an open postoperative wound; right, a representative augmented derivative produced by the manual compositing pipeline (feature pasting, cross-feature mixing, background substitution, and morphological operations).
Ground-truth image-level annotations were produced by 2 clinicians, a licensed physiotherapist and a fellowship-trained orthopedic surgeon, using a standardized label set comprising 6 categories, including infection, blood presence, purulent discharge, open wound, discharge or exudate, and cellulitis. Labels were applied at the surgical-site level to each EO image. To ensure diagnostic accuracy, all infection labels were cross-referenced against contemporaneous clinical records, including physical examination findings and serological markers of infection; where clinical record findings conflicted with the image-level annotation, the clinical record superseded the annotator label, and the annotation was updated accordingly. All annotations were reviewed for completeness and consistency by the principal orthopedic investigator prior to inclusion in the training dataset. Wound pixel-level segmentation and anatomical landmark labeling were performed separately by the research team under the supervision of the principal orthopedic investigator.
All statistical tests were 2-tailed with α=.05. Continuous variables were assessed for normality with the Shapiro-Wilk test and summarized as mean (SD) and 95% CI or, when nonnormal, as median (IQR); proportions are reported with numerators, denominators, and Wilson 95% CIs.
Two-group comparisons of continuous variables used the Mann-Whitney
All performance metrics were computed on the held-out sets described under dataset and development sets. For the thermal classifier, discrimination was summarized by the area under the receiver operating characteristic curve (AUROC), and, given class imbalance, operating points were reported in clinical terms: sensitivity, specificity, and positive and negative predictive value, and the number needed to evaluate (defined as 1/positive predictive value). Single-figure operating points were obtained by nested threshold selection, with the decision threshold chosen on training patients and applied to held-out patients. CIs for classifier metrics were 95% percentile bootstrap intervals from patient-stratified bootstrap resampling (2000 resamples, preserving prevalence). Segmentation performance was summarized by the dice coefficient and intersection-over-union, reported per class and as macro- and micro-averages; detection performance by mean average precision under the Common Objects in Context (COCO) protocol; and region-of-interest segmentation by intersection over union. Segmentation and detection CIs were 95% percentile bootstrap intervals over images (1000 resamples). Machine learning–native measures are reported in full in
Statistical analyses were performed in Python 3.12 (
We developed a sensor-agnostic pipeline that fuses EO imaging with infrared radiometry to convert postoperative data into structured, clinically interpretable indicators of wound status (
Detections from the EO stream are transferred to thermal space by a deterministic EO-to-infrared projection that preserves geometry and restricts analysis to clinically relevant zones. Within these mapped regions, an infrared segmentation model isolates the radiometric ROI, suppressing environmental background while preserving landmark context. Finally, a radiometric classifier analyzes the thermal data and returns patient-level risk scores and spatial risk maps; its operating threshold can be selected out of sample to match clinical workload and risk tolerance, from a high-sensitivity rule-out setting to a high-specificity rule-in setting. EO models were developed on expert-annotated institutional images augmented with publicly available data and evaluated on held-out sets, and the thermal classifier was evaluated by patient-grouped cross-validation.
Architecture of the multimodal electro-optical (EO) and infrared analysis pipeline. EO: electro-optical; IR: infrared; RGB: red, green, and blue; ROI: region of interest.