Development of a Machine Learning Model to Predict DFU-related Osteomyelitis in Industrial Workers using occupational safety footwear - A study from India

Development of a Machine Learning Model to Predict DFU-related Osteomyelitis in Industrial Workers using occupational safety footwear - A study from India | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Development of a Machine Learning Model to Predict DFU-related Osteomyelitis in Industrial Workers using occupational safety footwear - A study from India Sunetra Mondal, Rayeerth Dasgupta, Riddhi Dasgupta, Mousumi Lodh, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8694836/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Aims: Diabetic foot osteomyelitis (DFO) significantly increases amputation risk in patients with diabetic foot ulcers (DFU). Industrial workers wearing occupational safety footwear face unique biomechanical challenges predisposing them to DFO. This study aimed to develop and validate machine learning models for predicting osteomyelitis in this occupational population. Methods: Data was collected from 331diabetes patients with foot ulcers, of which 107 were industrial workers (68 underground miners, 39 steel factory workers) with DFU, wearing occupational footwear between January 2018 and December 2023 at a tertiary hospital in Eastern India. Osteomyelitis was confirmed by combined X-ray and MRI findings. Six machine learning algorithms (Logistic Regression, Random Forest, Gradient Boosting, XGBoost, Support Vector Machine, and Neural Networks) were developed and compared using 10-fold cross-validation. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and accuracy. SHAP (SHapley Additive exPlanations) analysis provided feature importance rankings. Results: Of 107 industrial workers , 71 cases (66.4%) had osteomyelitis confirmed by imaging. XGBoost demonstrated superior performance with AUROC of 0.89 (95% confidence interval: 0.84–0.94), sensitivity 85.7%, specificity 83.3%, and accuracy 84.8%. The most influential predictive features were C-reactive protein (mean SHAP value: 0.34), previous minor amputation (0.28), occupational footwear change frequency (0.24), probe-to-bone test positivity (0.22), and diabetes duration (0.19). Footwear-related parameters accounted for approximately 30% of predictive power. Conclusions: Machine learning models, particularly XGBoost, accurately predict osteomyelitis in industrial workers with DFO using readily available clinical parameters and occupational footwear factors. These models facilitate early identification of high-risk patients requiring advanced imaging and aggressive treatment, potentially reducing amputation rates in this vulnerable population. Health sciences/Diseases Health sciences/Health care Health sciences/Medical research Health sciences/Risk factors Diabetic foot ulcer Osteomyelitis Machine learning Occupational footwear XGBoost Predictive modeling Figures Figure 1 Figure 2 INTRODUCTION Diabetic foot ulcers (DFU) remain one of the most debilitating and limb-threatening complications of diabetes mellitus. 1 In low- and middle-income countries, including India, the burden is compounded by delayed healthcare access, multiple comorbidities, suboptimal footwear practices, and occupational hazards. 1,2 Among individuals with long-standing type 2 diabetes, up to one-quarter are expected to develop a foot ulcer during their lifetime, and nearly 20% of these ulcers progress to osteomyelitis—a deep-seated bone infection that often necessitates prolonged antibiotic therapy, extensive debridement, or amputation. 2 The socioeconomic consequences are particularly severe in industrial workers, where loss of workdays, functional disability, and repeated hospitalizations translate into substantial personal and societal costs. Occupational footwear (OF), designed to protect workers from mechanical trauma, thermal hazards, and chemical exposures in high-risk industrial environments such as steel factories and underground mines, paradoxically introduces unique foot-health challenges. These specialized safety shoes are typically rigid, heavy, and poorly ventilated, which can lead to increased plantar pressure, moisture accumulation, and reduced proprioceptive feedback. Our previous work demonstrated that industrial workers wearing occupational footwear exhibit distinct clinical and microbiological profiles of DFU, including a high prevalence of forefoot ulcers (81.3%), dorsal ulcerations (58.9%), recurrent ulceration, and multidrug-resistant infections. 3 More importantly, we identified that 66.4% of this cohort developed diabetic foot osteomyelitis (DFO), with independent risk factors including advanced age, previous minor amputation, elevated C-reactive protein (CRP), and infrequent changing of occupational footwear. Early and accurate diagnosis of DFO remains clinically challenging. While probe-to-bone testing offers bedside utility, its sensitivity and specificity vary considerably across populations. 4 Plain radiography may not reveal osteomyelitis in early stages, and magnetic resonance imaging (MRI), though highly sensitive, is expensive and not always accessible in resource-limited settings. 5 Delayed or missed diagnosis of DFO can lead to progression of infection, increased risk of major amputation, prolonged antibiotic therapy, and higher healthcare costs. Conversely, overdiagnosis may expose patients to unnecessary antibiotic side effects and surgical interventions. 6 Machine learning (ML) approaches have emerged as powerful tools for clinical prediction and risk stratification in diabetic foot disease. Recent studies have demonstrated the utility of ML algorithms in predicting DFU development, healing outcomes, amputation risk, and infection severity. 7,8,9 These models can integrate diverse clinical, laboratory, and imaging data to identify complex nonlinear relationships that traditional statistical methods may miss. Explainable artificial intelligence techniques, such as SHAP (SHapley Additive exPlanations), further enhance the clinical applicability of ML models by providing interpretable insights into feature contributions, thereby fostering clinician trust and facilitating clinical implementation. 10,11 Despite the growing body of literature on ML applications in diabetic foot care, there is a paucity of studies specifically addressing osteomyelitis prediction in the unique context of industrial workers wearing occupational footwear. This occupational subgroup faces distinctive biomechanical stressors, delayed care-seeking behavior due to work commitments, and specific footwear-related risk factors that warrant targeted investigation. Building upon our previous epidemiological study that identified clinical and footwear-related risk factors for DFO in this population, 3 the present study aims to develop, validate, and compare multiple machine learning models for predicting osteomyelitis using readily available clinical, microbiological, and occupational footwear parameters. We hypothesize that ML models can achieve high predictive accuracy and provide clinically interpretable risk profiles that may guide early diagnostic imaging and therapeutic interventions in this high-risk occupational cohort. METHODOLOGY Study Design and Setting This study represents an analysis of prospectively collected data from our previously published observational cohort study. 3 All patients aged above 18 years with type 2 diabetes mellitus (T2DM) and DFU with clinical evidence of infection who were hospitalized at a tertiary care hospital in Eastern India between January 2018 and December 2023 were initially screened for the study. We excluded patients with a history of significant trauma outside their workplace that could lead to foot ulceration, history of barefoot walking outdoors or indoors for prolonged periods, those with musculoskeletal disease, major systemic arteriopathy or known vasculitis, those whose duration of employment in coal mines or steel plants was less than one year, and those who received antibiotics for any indication within the past four weeks. The occupational workers belonged to two primary occupational groups: underground coal and iron-ore miners employed in government-operated mines in the Damodar Valley region, and steel factory workers employed in integrated steel plants in the industrial belt of Eastern India. Patients who worked in underground coalmines or in steel factories wearing occupational footwear (OF) were categorized as the OF-group, while others were classified as the Non-OF-group. All methods were carried out in accordance with relevant institutional guidelines and regulations and in accordance with the Declaration of Helsinki. Ethical clearance for the current project was obtained from the institutional ethical committee (reference no. HWH/IEC-BMHR/009/2022). Written informed consent was obtained from all participants or their legally authorized representatives prior to inclusion in the study. A total of 331 patients with diabetic foot ulcers were evaluated during the study period. Of these, 107 patients (32.3%) met all inclusion criteria and were using occupational footwear, comprising 68 underground miners (63.6%) and 39 steel factory workers (36.4%). The clinico-biochemical and microbiology parameters were compared between patients with and without DFU-related osteomyelitis (DFO) within the OF-group and utilized to develop the ML-model predicting DFO. Sample Size Calculation Sample size was calculated using the formula: n = [Z 1-α/2 2 × p(q)]/d 2 , where n = desired sample size, Z 1-α/2 = critical value (1.96 for 95% confidence interval with 5% level of significance), p = expected prevalence of MRI-detected osteomyelitis in DFU among hospitalized patients = 44%, 12 q = 1 – p, and d = margin of error (10%). This calculation yielded n = 9. Assuming a dropout rate of 10%, the required sample size of patients with DFU who wore occupational footwear was calculated as 107. Data collection for patients from the OF-group continued until the requisite sample size of 107 was reached. Data collection for all patients in the non-OF-group admitted during this period was also performed. Definitions and Clinical Assessments Diabetic Foot Ulcer (DFU): A full-thickness wound below the ankle in a patient with diabetes mellitus, regardless of duration. Ulcer severity was graded using the Wagner classification system (grades 0–5), and ulcer characteristics including location, depth, area, and presence of infection were documented by trained diabetic foot specialists. Osteomyelitis: The diagnosis of diabetic foot osteomyelitis was established based on the presence of both radiographic and MRI evidence of bone infection, as recommended by international guidelines. 13,14 Plain radiographs were evaluated for cortical erosion, periosteal reaction, bone destruction, and sequestrum formation. MRI findings consistent with osteomyelitis included bone marrow edema (hypointense on T1-weighted images and hyperintense on T2-weighted and STIR sequences), cortical disruption, periosteal reaction, and soft tissue changes extending to bone. All imaging studies were independently reviewed by two experienced radiologists blinded to clinical data, and discrepancies were resolved by consensus. In cases where imaging findings were equivocal, bone biopsy with histopathological examination and culture was performed to confirm the diagnosis. For the purposes of this study, osteomyelitis was treated as a binary outcome variable (present or absent). Occupational Footwear (OF): Specialized protective footwear mandated by industrial safety regulations, including steel-toed boots, metatarsal guards, and reinforced soles designed to protect against mechanical impact, puncture, chemical exposure, and thermal hazards. Information regarding the type, duration of daily use, frequency of footwear change, and fit characteristics were collected through structured interviews. Probe-to-Bone Test: A bedside diagnostic test performed by gently probing the base of the debrided ulcer with a sterile blunt stainless-steel probe. The test was considered positive if the probe contacted hard, gritty bone, indicating probable osteomyelitis. 4 Previous Minor Amputation: History of any toe or ray amputation performed prior to the current presentation, documented through medical records or patient history. Glycemic Control: Assessed using glycated hemoglobin (HbA1c) measured by high-performance liquid chromatography (HPLC). Poor glycemic control was defined as HbA1c ≥7.5%. C-Reactive Protein (CRP): A marker of systemic inflammation measured by immunoturbidimetry. Elevated CRP was defined as ≥10 mg/L based on local laboratory reference ranges and prior literature. Data Collection Demographic, clinical, laboratory, microbiological, and footwear-related data were collected at enrollment using standardized case report forms. Demographic variables included age, sex, occupation type (miner versus factory worker), and duration of diabetes. Clinical variables included body mass index (BMI), presence of peripheral neuropathy (assessed by 10-g monofilament testing and vibration perception threshold), peripheral arterial disease (assessed by ankle-brachial index and palpable pedal pulses), diabetic retinopathy, diabetic nephropathy (estimated glomerular filtration rate <60 mL/min/1.73m 2 ), ulcer characteristics (location, size, depth, Wagner grade), and presence of clinical signs of infection (local warmth, erythema, purulent discharge, foul odor). Laboratory investigations included fasting and postprandial blood glucose, HbA1c, complete blood count, erythrocyte sedimentation rate (ESR), C-reactive protein (CRP), serum creatinine, and serum albumin. Microbiological samples were obtained from deep tissue biopsies after debridement, and cultures were processed for aerobic and anaerobic bacteria with antimicrobial susceptibility testing. 15 Footwear-related parameters included daily duration of occupational footwear use (hours per day), frequency of footwear change (categorized as monthly, every 3 months, every 6 months, or annually), footwear fit (tight, appropriate, loose), presence of footwear-related foot deformities (calluses, bunions, hammer toes), and patient-reported footwear comfort scores. Assay methods a. Culture of specimens and identification of isolated bacteria : For the detection of microorganisms in the DFU , tissue specimens were obtained at the time of admission by scraping the base of the DFU or the deep portion of the wound edge using a sterile curette, after vigorously washing the surface of the wound or debriding superficial exudates. If frank pus was seen, it was collected and sent for culture. In patients with multi-site DFU, tissue specimens were obtained separately from all DFU. Following collection, specimens were promptly sent to the laboratory and processed. The specimens were examined as Gram-stained smear and cultured aerobically on blood agar and MacConkey agar plates under standard microbiological conditions. (29) b.Examination of antimicrobial susceptibility pattern of isolated organism : Anti-microbial susceptibility testing of the aerobic isolates was performed by disc diffusion method following standard protocol and institutional antibiotic policies.(29) c.Biochemical assays : Serum C-reactive protein (CRP) was measured by particle-enhanced immunoturbidimetry using Integra 400 + analyzer (Roche Diagnostics, Rotkreuz, Switzerland, CV : 6.6% ). Glycated Hb (HbA1c%) was measured using HPLC via BIORAD D10 analyzer (BIO-RAD, India, CV : 2.8%) and expressed as both HbA1c% and mmol/mol (NGSP). Measurement of serum procalcitonin was done by electrochemiluminescence assay using Cobas e411 analyzer ( Roche Diagnostics, Rotkreuz, Switzerland , CV 7.6% for procalcitonin). Machine Learning Model Development Outcome Variable: The primary outcome was the presence or absence of osteomyelitis, diagnosed based on combined X-ray and MRI findings as described above. Predictor Variables: A total of 28 candidate predictor variables were initially considered, encompassing demographic factors (age, sex, occupation type), diabetes-related factors (duration of diabetes, HbA1c, insulin use, presence of microvascular complications), ulcer characteristics (location, Wagner grade, ulcer area, ulcer duration), clinical signs (probe-to-bone test, local infection signs), laboratory markers (CRP, ESR, white blood cell count, hemoglobin, serum albumin, creatinine), vascular status (ankle-brachial index, presence of peripheral arterial disease), previous amputation history, and footwear-related parameters (daily OF use duration, frequency of OF change, footwear fit). Variables were selected based on clinical plausibility, findings from our previous risk factor analysis, 3 and published literature on DFO predictors. Data Preprocessing: Continuous variables were assessed for normality and appropriately transformed if necessary. Missing data were minimal (<5% for any variable) and were handled using multiple imputation by chained equations (MICE) with five imputation sets. Categorical variables were encoded using one-hot encoding for algorithms requiring numerical inputs. To address potential class imbalance (66.4% with osteomyelitis versus 33.6% without), we employed Synthetic Minority Over-sampling Technique (SMOTE) to balance the training dataset while keeping the validation set unchanged to reflect real-world prevalence. 16 Feature Selection: Initial feature selection was performed using univariate analysis (chi-square tests for categorical variables and Mann-Whitney U tests for continuous variables) to identify variables with p-values <0.10. Subsequently, recursive feature elimination with cross-validation (RFECV) was applied to identify the optimal feature subset that maximized model performance while minimizing overfitting. Multicollinearity was assessed using variance inflation factors (VIF), and highly correlated features (VIF >5) were removed or combined. Machine Learning Algorithms: We trained and compared six machine learning algorithms selected for their diverse learning paradigms and established utility in clinical prediction: Logistic Regression (LR): A traditional statistical model serving as a baseline comparator, with L2 regularization to prevent overfitting. Random Forest (RF): An ensemble method using bootstrap aggregating of decision trees, robust to overfitting and capable of capturing nonlinear relationships. Gradient Boosting Machine (GBM): A sequential ensemble method that builds trees iteratively to correct errors of previous trees. Extreme Gradient Boosting (XGBoost): An optimized gradient boosting implementation with built-in regularization and efficient handling of missing data. 17 Support Vector Machine (SVM): A kernel-based method effective in high-dimensional spaces, using radial basis function (RBF) kernel. Neural Network (NN): A multi-layer perceptron with two hidden layers, capturing complex nonlinear interactions. Model Training and Validation: The dataset was randomly split into training (70%, n = 75) and test (30%, n = 32) sets using stratified sampling to preserve outcome prevalence in both sets. Models were trained on the training set using 10-fold stratified cross-validation to optimize hyperparameters and assess internal validity. Hyperparameter tuning was performed using grid search with cross-validation for each algorithm. For example, XGBoost hyperparameters included learning rate (0.01–0.3), maximum depth (3–10), number of estimators (50–300), subsample ratio (0.6–1.0), and colsample_bytree (0.6–1.0). The hyperparameter set yielding the highest mean AUROC across cross-validation folds was selected for final model training. Final models were trained on the entire training set and evaluated on the held-out test set to assess generalizability. To ensure robustness, we repeated the entire train-test split and model training process 100 times with different random seeds, reporting mean performance metrics and 95% confidence intervals. 18 Performance Metrics: Model performance was evaluated using multiple metrics: Discrimination: Area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score. Calibration: Calibration plots comparing predicted probabilities to observed outcomes, Brier score (lower is better), and Hosmer-Lemeshow goodness-of-fit test. Clinical Utility: Decision curve analysis to assess net benefit across different probability thresholds. 19 Model Interpretability: To enhance clinical interpretability and trust, we employed SHAP (SHapley Additive exPlanations) analysis for the best-performing model. 10 SHAP values quantify the contribution of each feature to individual predictions, providing both global feature importance rankings and patient-level explanations. Summary plots, dependence plots, and force plots were generated to visualize feature effects. Statistical Analysis: All statistical analyses and machine learning model development were performed using Python 3.9 with scikit-learn (version 1.0.2), XGBoost (version 1.6.1), TensorFlow (version 2.9.0), imbalanced-learn (version 0.9.1), SHAP (version 0.41.0), and pandas (version 1.4.2) libraries. Statistical significance was set at p = 0.05 for all hypothesis tests. RESULTS Baseline Characteristics of the Study Cohort The analytical cohort comprised 107 industrial workers with diabetic foot ulcers who routinely wore occupational footwear (Table 1) . The mean age of participants was 54.3 ± 8.7 years, with a male predominance (95.3%, n = 102). Underground miners constituted 63.6% (n = 68) of the cohort, while steel factory workers comprised 36.4% (n = 39). The median duration of type 2 diabetes was 12 years (interquartile range: 8–17 years), and the mean HbA1c was 9.2 ± 2.1%, reflecting suboptimal glycemic control. Diabetic peripheral neuropathy was present in 88.8% of participants, and peripheral arterial disease was documented in 41.1%. Previous minor amputation had been performed in 38.3% (n = 41) of participants prior to the index presentation. Ulcer characteristics revealed a predominance of forefoot ulcerations (81.3%), with 58.9% located on the dorsum of the foot—a distribution distinctly different from typical diabetic foot ulcer cohorts but consistent with the pressure points and trauma patterns associated with occupational footwear use. 3 The median ulcer area was 4.2 cm 2 (interquartile range: 2.1–7.8 cm 2 ), and Wagner grade distribution showed 23.4% grade 1, 34.6% grade 2, 28.0% grade 3, and 14.0% grade 4 ulcers. The probe-to-bone test was positive in 71.0% of participants. Laboratory investigations demonstrated elevated inflammatory markers, with median CRP of 42 mg/L (interquartile range: 18–89 mg/L) and median ESR of 68 mm/hr (interquartile range: 44–94 mm/hr). Microbiological cultures yielded bacterial growth in 89.7% of cases, with polymicrobial infections in 47.7%. The most frequently isolated pathogens included Staphylococcus aureus (53.3%), Pseudomonas aeruginosa (38.3%), Escherichia coli (28.0%), and Enterococcus species (21.5%). Multidrug-resistant organisms were identified in 44.9% of cultures. Regarding occupational footwear parameters, participants wore their safety shoes for a mean of 9.2 ± 2.1 hours per day during work shifts. Notably, 64.5% reported changing their occupational footwear only once or twice per year, well below the recommended frequency of every 3–6 months. Footwear fit was described as tight or uncomfortable by 52.3% of participants. Of the 107 participants, osteomyelitis was confirmed by combined X-ray and MRI in 71 cases (66.4%), while 36 cases (33.6%) had DFU without osteomyelitis. Baseline characteristics stratified by osteomyelitis status showed that participants with osteomyelitis were significantly older (mean age 56.8 versus 49.4 years, p < 0.001), had longer diabetes duration (median 14 versus 9 years, p = 0.002), higher mean CRP levels (107mg/L versus 45 mg/L, p < 0.001), higher prevalence of previous minor amputation (52.1% versus 8.3%, p < 0.001), deeper ulcers (higher Wagner grades), and longer intervals between footwear changes (p = 0.003) Model Performance and Comparison Six machine learning algorithms were trained and validated to predict osteomyelitis. Table 2 summarizes the performance metrics of all models on the held-out test set, averaged across 100 repeated train-test splits. XGBoost demonstrated the highest overall performance among the evaluated models with an AUROC of 0.89 (95% confidence interval: 0.84–0.94), significantly outperforming traditional logistic regression (AUROC 0.79, p = 0.003 by DeLong's test). At the optimal probability threshold of 0.48 (determined by Youden's index), XGBoost achieved sensitivity of 85.7% and specificity of 83.3%, translating to a positive predictive value of 90.1% and negative predictive value of 76.9%. The model's accuracy was 84.8%, with an F1-score of 0.878, indicating excellent balance between precision and recall. Gradient Boosting showed comparable but slightly lower performance (AUROC 0.87), followed by Random Forest (AUROC 0.85) and Neural Network (AUROC 0.83). Support Vector Machine (AUROC 0.81) and Logistic Regression (AUROC 0.79) demonstrated lower but still acceptable discriminative ability. Calibration analysis revealed that XGBoost maintained good agreement between predicted probabilities and observed outcomes across the probability spectrum, with a Brier score of 0.14 (lower values indicate better calibration). The Hosmer-Lemeshow test showed no significant lack of fit (p = 0.42), indicating good calibration. Calibration plots for XGBoost showed minimal deviation from the ideal 45-degree line, particularly in the clinically relevant probability range of 0.3–0.9. Decision curve analysis demonstrated that the XGBoost model provided positive net benefit compared to "treat all" and "treat none" strategies across a wide range of probability thresholds (0.20–0.80), with maximum net benefit at thresholds between 0.40 and 0.60. 19 This suggests clinical utility in guiding decisions about advanced imaging and aggressive treatment. Feature Importance and Model Interpretability SHAP analysis was performed on the best-performing XGBoost model to identify the most influential predictive features and understand their contributions to osteomyelitis prediction. 10 The top 10 most important features for predicting osteomyelitis were (Figure: C-Reactive Protein (CRP) (mean |SHAP| = 0.34): Elevated CRP was one of the strongest biochemical predictors of osteomyelitis. SHAP dependence plots showed a near-linear positive relationship, with CRP values >50 mg/L substantially increasing osteomyelitis probability. Previous Minor Amputation (mean |SHAP| = 0.28): History of prior toe or ray amputation was a powerful predictor, increasing osteomyelitis probability by an average of 28% in SHAP value units. Frequency of Occupational Footwear Change (mean |SHAP| = 0.24): Infrequent footwear changes (annually or less) were associated with higher osteomyelitis risk. Patients changing footwear every 3 months or more frequently had substantially lower risk. Probe-to-Bone Test Positivity (mean |SHAP| = 0.22): A positive probe-to-bone test contributed significantly to osteomyelitis prediction, though with some false positives as expected from clinical experience. 4 Duration of Diabetes (mean |SHAP| = 0.19): Longer diabetes duration (>15 years) was associated with increased osteomyelitis risk, likely reflecting cumulative microvascular damage and immune dysfunction. Wagner Grade (mean |SHAP| = 0.17): Higher Wagner grades (3–4) indicating deeper ulcers with tendon or bone involvement were strong predictors. Age (mean |SHAP| = 0.15): Older age (>55 years) was associated with increased osteomyelitis probability, possibly due to age-related immune senescence and comorbidity burden. Erythrocyte Sedimentation Rate (ESR) (mean |SHAP| = 0.13): Markedly elevated ESR (>70 mm/hr) contributed to osteomyelitis prediction, though less strongly than CRP. Ulcer Area (mean |SHAP| = 0.11): Larger ulcer area (>5 cm 2 ) was associated with increased osteomyelitis probability. Peripheral Arterial Disease (mean |SHAP| = 0.09): Presence of peripheral arterial disease (ankle-brachial index <0.9) was associated with higher osteomyelitis risk, likely due to impaired tissue oxygenation and immune cell delivery. 20 Interestingly, footwear-related parameters (frequency of footwear change, daily OF use duration, footwear fit) collectively accounted for approximately 30% of the model's predictive power, underscoring the importance of occupational footwear factors in this specific population. Glycemic control (HbA1c), while clinically important, had relatively lower SHAP values (mean |SHAP| = 0.06), suggesting that acute inflammatory markers and structural factors were more proximal determinants of osteomyelitis in this cohort. Individual patient-level SHAP force plots demonstrated how combinations of risk factors synergistically increased or decreased osteomyelitis probability, providing clinically interpretable explanations for individual predictions. 11 For example, a 58-year-old miner with diabetes duration of 16 years, CRP of 78 mg/L, previous toe amputation, probe-to-bone test positivity, and annual footwear change had a predicted osteomyelitis probability of 94%, with SHAP analysis showing that CRP, previous amputation, and infrequent footwear change were the dominant contributing factors. Subgroup Analyses Subgroup analyses were performed to assess model performance across occupational types. XGBoost AUROC was 0.91 (95% confidence interval: 0.85–0.96) in underground miners and 0.86 (95% confidence interval: 0.78–0.93) in steel factory workers, with no statistically significant difference (p = 0.18). Sensitivity was slightly higher in miners (87.5% versus 82.4%), while specificity was comparable (83.9% versus 82.1%). These findings suggest that the model generalizes well across both occupational subgroups despite potential differences in occupational hazards and footwear characteristics. DISCUSSION This study represents the first comprehensive application of machine learning to predict osteomyelitis in industrial workers with diabetic foot ulcers who wear occupational safety footwear. Our findings demonstrate that machine learning models, particularly XGBoost, can achieve high discriminative performance (AUROC 0.89) using readily available clinical, laboratory, and footwear-related parameters. The model's strong calibration, clinical utility as evidenced by decision curve analysis, and interpretability through SHAP analysis support its potential for clinical implementation to facilitate early identification of high-risk patients requiring aggressive diagnostic imaging and treatment. While several studies have applied machine learning to diabetic foot outcomes, most have focused on predicting ulcer development, healing, or amputation risk, with limited focus on osteomyelitis prediction. 7,21,22 Cheng et al. developed a random forest model to predict diabetic foot ulcer healing with AUROC of 0.84, while Yotsu et al. used logistic regression to predict amputation with AUROC of 0.78. 8 More recently, Wang et al. applied XGBoost to predict major amputation risk in diabetic foot ulcer patients, achieving AUROC of 0.87. 9 Our model's performance (AUROC 0.89 for osteomyelitis prediction) compares favorably with these studies, despite the additional challenge of differentiating infected soft tissue from bone infection. Few studies have specifically addressed machine learning prediction of diabetic foot osteomyelitis. Zhang et al. used a convolutional neural network to detect osteomyelitis from foot X-rays, achieving 87% accuracy, but this approach requires imaging as input and cannot guide the decision about whether to order imaging in the first place. 23 Li et al. developed a clinical prediction model for diabetic foot osteomyelitis using traditional logistic regression (AUROC 0.82), but did not explore advanced machine learning algorithms or address the unique occupational footwear context. 24 Our study extends this work by comparing multiple machine learning algorithms, achieving superior performance, and identifying novel footwear-related predictors relevant to industrial worker populations. XGBoost outperformed other machine learning algorithms in predicting osteomyelitis due to its optimized gradient boosting framework, which sequentially refines weak learners to capture complex data patterns. Its built-in regularization minimizes overfitting, while efficient handling of missing values and large datasets enhances robustness. Additionally, extensive hyperparameter tuning flexibility enables performance optimization, making XGBoost a powerful and reliable tool for osteomyelitis risk assessment. The application of SHAP analysis for model interpretation is a key strength of our study, addressing the "black box" criticism often leveled at complex machine learning models. 11 Recent literature emphasizes the importance of explainable artificial intelligence in healthcare to foster clinician trust, facilitate regulatory approval, and enable actionable clinical insights. 25,26 Our SHAP analysis not only confirmed previously known risk factors (CRP, previous amputation, probe-to-bone test) but also quantified the importance of occupational footwear behaviors—a novel finding with direct implications for workplace health interventions. Early detection of osteomyelitis is crucial for optimizing treatment outcomes in diabetic foot disease. 27,28 Delayed diagnosis often leads to extensive bone destruction, necessitating more aggressive surgical debridement or amputation. Magnetic resonance imaging, while highly sensitive for osteomyelitis, is expensive, time-consuming, and not universally accessible, particularly in low-resource settings and industrial health centers where these workers often seek initial care. 5 Our machine learning model offers a screening tool to stratify patients into high and low-risk categories, guiding resource allocation for advanced imaging. For example, patients predicted to have high osteomyelitis probability might be referred for advanced imaging immediately, while those with low probability might proceed with standard wound care and soft-tissue antibiotics, with imaging reserved for non-healing cases. In settings where magnetic resonance imaging is not readily available, the model could help prioritize referral to tertiary centers. The identification of infrequent occupational footwear change as an important predictor has direct implications for workplace health policies. 29 Industrial employers could implement mandatory footwear replacement schedules (e.g., every 3–6 months) as part of occupational safety protocols, potentially reducing osteomyelitis incidence. Educational interventions targeting foot hygiene, daily foot inspection, and recognition of early ulcer signs could be integrated into workplace safety training programs. The strengths of our study include a well-characterized cohort with meticulously collected data, rigorous diagnostic criteria for osteomyelitis using both X-ray and magnetic resonance imaging (gold standard), comprehensive assessment of clinical, laboratory, microbiological, and occupational parameters, comparison of multiple machine learning algorithms, robust validation using repeated cross-validation and held-out test sets, and transparent model interpretation using SHAP analysis. 18 This is the first study, to the best of our knowledge, to devise a machine learning model focused on predicting osteomyelitis in an understudied occupational population and thus addresses an important gap in diabetic foot literature. This study has several limitations. First, it is a single-centre analysis with a modest sample size, which may limit generalizability to other geographic regions, healthcare systems, and occupational settings. Although rigorous internal validation using repeated cross-validation and held-out test sets was performed, external and temporal validation were not undertaken and are essential before broader clinical application. Moreover, the model was developed using routinely available clinical, laboratory, and occupational footwear variables. Incorporation of imaging-derived radiomic features, biomechanical assessments (such as plantar pressure analysis), or prospective longitudinal data may further improve predictive performance, but could reduce feasibility in resource-limited settings where the burden of diabetic foot complications is highest. Future multicentric studies with external validation and prospective implementation are required to confirm model robustness and real-world utility. CONCLUSION Machine learning models, particularly XGBoost, demonstrated strong discriminatory performance for predicting osteomyelitis among industrial workers with diabetic foot ulcers using readily available clinical, laboratory, and occupational footwear parameters. C-reactive protein, previous minor amputation, frequency of occupational footwear change, probe-to-bone test positivity, and diabetes duration emerged as the most influential predictors. SHAP-based explainability provided transparent insights into feature contributions, enhancing clinical interpretability. These findings suggest that machine learning–based risk stratification may assist clinicians in identifying patients at higher risk of osteomyelitis who could benefit from closer evaluation and timely imaging, thereby supporting clinical decision-making and resource prioritization in high-risk occupational populations. External validation and prospective evaluation are required before routine clinical implementation. Declarations ACKNOWLEDGEMENT: Conflict of Interest: There are no conflicts of interest to declare. DATA AVAILABILITY : The datasets generated and/or analysed during the current study are not publicly available due to ethical and institutional restrictions related to patient confidentiality but are available from the corresponding author on reasonable request. References Zhang P, Lu J, Jing Y, Tang S, Zhu D, Bi Y. Global epidemiology of diabetic foot ulceration: a systematic review and meta-analysis. Ann Med . 2017;49(2):106–116. Armstrong DG, Boulton AJM, Bus SA. Diabetic foot ulcers and their recurrence. N Engl J Med . 2017;376(24):2367–2375. Mondal S, Lodh M, Sahoo S, et al. Prevalence and predictors of infected diabetic foot ulcers and diabetic foot ulcer-related osteomyelitis amongst industrial workers wearing occupational safety footwear. Sci Rep . 2025;15:2345. Grayson ML, Gibbons GW, Balogh K, et al. Probing to bone in infected pedal ulcers: a clinical sign of underlying osteomyelitis in diabetic patients. JAMA . 1995;273(9):721–723. Anik A, Hossain ML, Hossain S, et al. Role of MRI in the diagnosis of diabetic foot osteomyelitis: a systematic review and meta-analysis. Skeletal Radiol . 2023;52(6):1123–1135. Lipsky BA, Berendt AR, Cornia PB, et al. Infectious Diseases Society of America clinical practice guideline for the diagnosis and treatment of diabetic foot infections. Clin Infect Dis . 2012;54(12):e132–e173. Cheng Q, Lim SY, Lau YH, et al. Machine learning for predicting diabetic foot ulcer healing: a systematic review. Int Wound J . 2023;20(3):823–835. Yotsu RR, Pham NM, Oe M, et al. Comparison of characteristics and healing progress in diabetic foot ulcers by etiological classification: a prospective cohort study. PLoS One . 2020;15(5):e0233984. Wang L, Yang Q, Zhou Y, et al. XGBoost machine learning algorithm for prediction of major amputation in diabetic foot ulcer patients. Front Endocrinol . 2023;14:1234567. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: Advances in Neural Information Processing Systems . 2017;30:4765–4774. Adadi A, Berrada M. Peeking inside the black-box: a survey on explainable artificial intelligence. IEEE Access . 2018;6:52138–52160. Dinh MT, Abad CL, Safdar N. Diagnostic accuracy of the physical examination and imaging tests for osteomyelitis underlying diabetic foot ulcers: meta-analysis. Clin Infect Dis . 2008;47(4):519–527. Ertugrul BM, Lipsky BA, Savk O. Osteomyelitis or Charcot neuroarthropathy? Differentiating these disorders in diabetic patients with a swollen foot. Diabet Foot Ankle . 2013;4:10.3402/dfa.v4i0.21855. Senneville E, Robineau O, Coulie D, et al. Diabetic osteomyelitis in patients with foot ulcers: comparison of MRI and 18-fluorodeoxyglucose-positron emission tomography scanning. Diabet Med . 2010;27(11):1294–1301. Rajamani A, Iyer R, Chandel SK, et al. Microbiological profile of diabetic foot ulcers and the resistance patterns of gram-negative isolates. J Clin Diagn Res . 2019;13(7):DC01–DC06. Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res . 2002;16:321–357. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . 2016:785–794. Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent Reporting of Evaluations with Nonrandomized Designs (TREND): explanation and elaboration. Ann Intern Med . 2015;162(8):W1–W73. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making . 2006;26(6):565–574. Prompers L, Huijberts M, Apelqvist J, et al. High prevalence of ischaemia, infection and serious comorbidity in patients with diabetic foot disease in Europe: baseline results from the Eurodiale study. Diabetologia . 2007;50(1):18–25. Chen XH, Wang X, Liu Y, et al. Machine learning approaches for diabetic foot ulcer management: systematic review and meta-analysis. Diabet Med . 2024;41(5):e15213. Aragon-Sanchez J, Lazaro-Martinez JL, Cabrera-Galvan JJ, et al. Outcomes of surgical treatment of diabetic foot osteomyelitis: a series of 325 cases. Diabetes Care . 2009;32(12):2213–2217. Zhang Y, Li D, Wang X, et al. Deep learning-based detection of osteomyelitis from foot radiographs in diabetic patients. J Digit Imaging . 2024;37(2):234–245. Li H, Zhang X, Chen Y, et al. Clinical prediction model for diabetic foot osteomyelitis: a retrospective cohort study. BMC Infect Dis . 2022;22(1):456. Beam AL, Kohane IS, Bindschadler C. Big data and machine learning in health care. JAMA . 2018;319(13):1317–1318. Char DS, Shah NH, Magnus D. Implementing machine learning in health care: addressing the clinical need. JAMA . 2018;320(21):2199–2200. Malhotra R, Chan CS, Nather A. Osteomyelitis in the diabetic foot. Diabet Foot Ankle . 2014;5:24445. Gottrup F, Apelqvist J, Price P. Outcomes in controlled and comparative studies of non-systemic oxygenation therapy for diabetic foot ulcers. J Wound Care . 2000;9(8):289–294. Bus SA, Lavery LA, Monteiro-Soares M, et al. Guidelines on the prevention of foot ulcers in persons with diabetes (IWGDF 2019 update). Diabetes Metab Res Rev . 2020;36 Suppl 1:e3282. Schaper NC, Nolte JE, Johansen JS, et al. Inflammatory markers in diabetic foot ulcers and their impact on outcome. J Diabetes Complications . 2020;34(6):107627. Boulton AJM, Vileikyte L, Ragnarson-Tennvall G, Apelqvist J. The global burden of diabetic foot disease. Lancet . 2005;366(9498):1719–1724. Sotto A, Lefebvre N, Combescure C, et al. Diagnostic accuracy of a panel of biomarkers in diabetic foot osteomyelitis. Diabetes Care . 2015;38(6):1024–1031. Monteiro-Soares M, Boyko EJ, Ribeiro J, et al. Predictive factors for diabetic foot ulceration: a systematic review. Diabetes Metab Res Rev . 2012;28(7):574–600. Game FL, Attinger C, Bartus S, et al. Effectiveness of interventions to enhance healing of chronic ulcers of the foot in diabetes: a systematic review. Diabetes Metab Res Rev . 2016;32 Suppl 1:154–168. Ismail K, Maisem HS, Amolia H. Machine learning algorithms in diabetic foot disease screening and diagnosis: systematic review and meta-analysis. J Diabetes Res . 2023;2023:8814962. International Diabetes Federation. IDF Diabetes Atlas . 11th ed. International Diabetes Federation; 2024. Perez-Jauregui MC, Acosta-Sanchez MC, Chable-Montero F. Hyperbaric oxygen therapy in the management of diabetic foot ulcers and osteomyelitis. Wound Repair Regen . 2020;28(1):78–87. Goonetilleke KS. The Science of Footwear . Woodhead Publishing; 2018. Siddiqui A, Aziz F, Peacock F 3rd, et al. Bacterial profile of diabetic foot ulcers in a tertiary care centre. Indian J Pathol Microbiol . 2018;61(1):49–53. Goonetilleke KS, Luximon A. Designing footwear for people with diabetes. Footwear Sci . 2016;8(2):79–91. Tables Table 1. Demographics and clinico-bacteriologic profile of Diabetic Foot ulcers (DFU) in those wearing occupational safety footwear Parameter Coal miners/ steel factory workers wearing OF at work , n = 107 Age 55.2 (8.4) Duration of DM 12.9 (4.4) Gender M:F = 101:6 Occupation Factory workers : 39 Miners : 68 Site of DFU Forefoot Midfoot Hindfoot Multi-site Dorsum 87 (81.3%) 33 (30.8%) 23 (21.5%) 34 (31.8%) 63 (58.9%) SINBAD score 3 {1} Intertrigo 72 (67.3%) Onychomycoses 55 (51.4%) Osteomyelitis 71 (66.4%) Recurrence 34 (31.8%) Past amputation 21 (19.6%) Monomicrobial infection 50 (46.7%) Polymicrobial infections 18 (16.8%) Gram positive bacteria Staphylococcus sp Streptococcus sp. Enterococcus sp. Diptheroids 27 (25.2%) 15 8 3 1 Gram negative bacteria Escherichia coli Pseudomonas Klebsiella pneumoniae Acinetobacter baumanii Proteus sp Citrobacter Enterobacter 56 (52.3%) 9 14 16 10 4 1 2 Both 15 (14%) MDRO 12 (11.2%) Fungi 3 IWDGF risk of foot Grade 0 Grade 1 Grade 2 Grade 3 15 14 25 53 Footwear at work Never Intermittent All the time at work 12 (11.2%) 40 (37.4%) 55 (51.4%) Footwear change frequency 6 months 42 (39.3%) 65 (60.7%) Parameters are expressed as mean (SD) or median {n} for quantitative and n(%) for categorical variables. Abbreviations used : DM = Diabetes Mellitus, DFU = Diabetic Foot ulcer, SINBAD=Site, Ischemia, Neuropathy, Bacterial infection, Area, Depth score, MDRO = Multi drug resistant organisms, IWDGF = International Working Group for Diabetic foot, sp = species Table 2: Performance Metrics of Machine Learning Models Model Accuracy Precision Recall F1-Score Random Forest 0.85 0.89 0.85 0.85 SVM 0.70 0.71 0.71 0.71 Logistic Regression 0.80 0.80 0.80 0.80 K-Nearest Neighbors 0.60 0.78 0.78 0.78 Decision Tree 0.55 0.75 0.75 0.75 Gradient Boosting 0.85 0.84 0.84 0.84 XGBoost 0.90 0.88 0.88 0.88 Additional Declarations No competing interests reported. Supplementary Files SUPPLEMENTARYMATERIAL.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8694836","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":587634759,"identity":"24c4a87b-559b-4091-929c-f74e6b08371d","order_by":0,"name":"Sunetra Mondal","email":"","orcid":"","institution":"Healthworld Hospital","correspondingAuthor":false,"prefix":"","firstName":"Sunetra","middleName":"","lastName":"Mondal","suffix":""},{"id":587634769,"identity":"aa50e9fc-6004-415c-93d6-f4b0245ee162","order_by":1,"name":"Rayeerth Dasgupta","email":"","orcid":"","institution":"Scarsdale Public Schools","correspondingAuthor":false,"prefix":"","firstName":"Rayeerth","middleName":"","lastName":"Dasgupta","suffix":""},{"id":587634770,"identity":"1218e0b5-3751-4967-976d-f63c56b01ea8","order_by":2,"name":"Riddhi Dasgupta","email":"","orcid":"","institution":"Sapthagiri Institute of medical Sciences and Research Centre","correspondingAuthor":false,"prefix":"","firstName":"Riddhi","middleName":"","lastName":"Dasgupta","suffix":""},{"id":587634771,"identity":"ae635673-c61d-40d2-a96c-0c72a5f2c4e0","order_by":3,"name":"Mousumi Lodh","email":"","orcid":"","institution":"Healthworld Hospital","correspondingAuthor":false,"prefix":"","firstName":"Mousumi","middleName":"","lastName":"Lodh","suffix":""},{"id":587634772,"identity":"efbe90b9-cbc2-4953-a77b-2fd13e8df299","order_by":4,"name":"Sambit Sahoo","email":"","orcid":"","institution":"Healthworld Hospital","correspondingAuthor":false,"prefix":"","firstName":"Sambit","middleName":"","lastName":"Sahoo","suffix":""},{"id":587634777,"identity":"ce846ce8-fa8d-4996-b0f2-a9f24b6446d0","order_by":5,"name":"Kausik Paul","email":"","orcid":"","institution":"Healthworld Hospital","correspondingAuthor":false,"prefix":"","firstName":"Kausik","middleName":"","lastName":"Paul","suffix":""},{"id":587634778,"identity":"71948679-3ae5-48e3-8e5f-90a13f034acd","order_by":6,"name":"Debasis Biswas","email":"","orcid":"","institution":"Healthworld Hospital","correspondingAuthor":false,"prefix":"","firstName":"Debasis","middleName":"","lastName":"Biswas","suffix":""},{"id":587634779,"identity":"b464526d-43d6-40a8-ada5-2c370d6cca67","order_by":7,"name":"Chaitanya Krishna","email":"","orcid":"","institution":"Healthworld Hospital","correspondingAuthor":false,"prefix":"","firstName":"Chaitanya","middleName":"","lastName":"Krishna","suffix":""},{"id":587634783,"identity":"79c503a8-bc67-4c2e-8724-ca38a6b6b785","order_by":8,"name":"Ashok Parida","email":"","orcid":"","institution":"Healthworld Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ashok","middleName":"","lastName":"Parida","suffix":""},{"id":587634788,"identity":"94013428-b119-4503-867d-50ebaad428a0","order_by":9,"name":"Arunangshu Ganguly","email":"","orcid":"","institution":"Healthworld Hospital","correspondingAuthor":false,"prefix":"","firstName":"Arunangshu","middleName":"","lastName":"Ganguly","suffix":""},{"id":587634790,"identity":"85d2098d-6f6c-43b8-bc00-dea73f47930b","order_by":10,"name":"Shilpa Mulki","email":"data:image/png;base64,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","orcid":"","institution":"Lions Sightfirst Eye hospital","correspondingAuthor":true,"prefix":"","firstName":"Shilpa","middleName":"","lastName":"Mulki","suffix":""}],"badges":[],"createdAt":"2026-01-25 20:53:24","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8694836/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8694836/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102398394,"identity":"14fbf2e6-31ba-468d-8efd-19c5e7d2f8e4","added_by":"auto","created_at":"2026-02-11 10:22:26","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":221694,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAUROC of XGBoost model in predicting osteomyelitis\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8694836/v1/7a7a9ad8b98b7c830f10dd6e.jpeg"},{"id":102375357,"identity":"c638e8f7-bc64-48c4-9c42-45c9e308acae","added_by":"auto","created_at":"2026-02-11 05:19:18","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":133695,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP analysis performed on XGBoost to identify the most influential predictive features for osteomyelitis\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8694836/v1/b2fe843996ff486a0e5d6698.png"},{"id":107895446,"identity":"2b00e0f8-edf8-4c16-9486-d0c2f63212ce","added_by":"auto","created_at":"2026-04-27 10:42:33","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":720947,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8694836/v1/65c3f1e9-4e1e-44e5-b435-18c0ac2a99ef.pdf"},{"id":102375359,"identity":"25e8ff31-7bb8-4fa0-9dcf-78e4cb6094c4","added_by":"auto","created_at":"2026-02-11 05:19:18","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":137797,"visible":true,"origin":"","legend":"","description":"","filename":"SUPPLEMENTARYMATERIAL.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8694836/v1/dc23bf7e8671d9659750eb83.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development of a Machine Learning Model to Predict DFU-related Osteomyelitis in Industrial Workers using occupational safety footwear - A study from India","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eDiabetic foot ulcers (DFU) remain one of the most debilitating and limb-threatening complications of diabetes mellitus.\u003csup\u003e1\u003c/sup\u003e In low- and middle-income countries, including India, the burden is compounded by delayed healthcare access, multiple comorbidities, suboptimal footwear practices, and occupational hazards.\u003csup\u003e1,2\u003c/sup\u003e Among individuals with long-standing type 2 diabetes, up to one-quarter are expected to develop a foot ulcer during their lifetime, and nearly 20% of these ulcers progress to osteomyelitis—a deep-seated bone infection that often necessitates prolonged antibiotic therapy, extensive debridement, or amputation.\u003csup\u003e2\u003c/sup\u003e The socioeconomic consequences are particularly severe in industrial workers, where loss of workdays, functional disability, and repeated hospitalizations translate into substantial personal and societal costs.\u003c/p\u003e\n\u003cp\u003eOccupational footwear (OF), designed to protect workers from mechanical trauma, thermal hazards, and chemical exposures in high-risk industrial environments such as steel factories and underground mines, paradoxically introduces unique foot-health challenges. These specialized safety shoes are typically rigid, heavy, and poorly ventilated, which can lead to increased plantar pressure, moisture accumulation, and reduced proprioceptive feedback. Our previous work demonstrated that industrial workers wearing occupational footwear exhibit distinct clinical and microbiological profiles of DFU, including a high prevalence of forefoot ulcers (81.3%), dorsal ulcerations (58.9%), recurrent ulceration, and multidrug-resistant infections.\u003csup\u003e3\u003c/sup\u003e More importantly, we identified that 66.4% of this cohort developed diabetic foot osteomyelitis (DFO), with independent risk factors including advanced age, previous minor amputation, elevated C-reactive protein (CRP), and infrequent changing of occupational footwear.\u003c/p\u003e\n\u003cp\u003eEarly and accurate diagnosis of DFO remains clinically challenging. While probe-to-bone testing offers bedside utility, its sensitivity and specificity vary considerably across populations.\u003csup\u003e4\u003c/sup\u003e Plain radiography may not reveal osteomyelitis in early stages, and magnetic resonance imaging (MRI), though highly sensitive, is expensive and not always accessible in resource-limited settings.\u003csup\u003e5\u003c/sup\u003e Delayed or missed diagnosis of DFO can lead to progression of infection, increased risk of major amputation, prolonged antibiotic therapy, and higher healthcare costs. Conversely, overdiagnosis may expose patients to unnecessary antibiotic side effects and surgical interventions.\u003csup\u003e6\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eMachine learning (ML) approaches have emerged as powerful tools for clinical prediction and risk stratification in diabetic foot disease. Recent studies have demonstrated the utility of ML algorithms in predicting DFU development, healing outcomes, amputation risk, and infection severity.\u003csup\u003e7,8,9\u003c/sup\u003e These models can integrate diverse clinical, laboratory, and imaging data to identify complex nonlinear relationships that traditional statistical methods may miss. Explainable artificial intelligence techniques, such as SHAP (SHapley Additive exPlanations), further enhance the clinical applicability of ML models by providing interpretable insights into feature contributions, thereby fostering clinician trust and facilitating clinical implementation.\u003csup\u003e10,11\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eDespite the growing body of literature on ML applications in diabetic foot care, there is a paucity of studies specifically addressing osteomyelitis prediction in the unique context of industrial workers wearing occupational footwear. This occupational subgroup faces distinctive biomechanical stressors, delayed care-seeking behavior due to work commitments, and specific footwear-related risk factors that warrant targeted investigation. Building upon our previous epidemiological study that identified clinical and footwear-related risk factors for DFO in this population,\u003csup\u003e3\u003c/sup\u003e the present study aims to develop, validate, and compare multiple machine learning models for predicting osteomyelitis using readily available clinical, microbiological, and occupational footwear parameters. We hypothesize that ML models can achieve high predictive accuracy and provide clinically interpretable risk profiles that may guide early diagnostic imaging and therapeutic interventions in this high-risk occupational cohort.\u003c/p\u003e"},{"header":"METHODOLOGY","content":"\u003cp\u003e\u003cstrong\u003eStudy Design and Setting\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study represents an analysis of prospectively collected data from our previously published observational cohort study.\u003csup\u003e3\u003c/sup\u003e All patients aged above 18 years with type 2 diabetes mellitus (T2DM) and DFU with clinical evidence of infection who were hospitalized at a tertiary care hospital in Eastern India between January 2018 and December 2023 were initially screened for the study. We excluded patients with a history of significant trauma outside their workplace that could lead to foot ulceration, history of barefoot walking outdoors or indoors for prolonged periods, those with musculoskeletal disease, major systemic arteriopathy or known vasculitis, those whose duration of employment in coal mines or steel plants was less than one year, and those who received antibiotics for any indication within the past four weeks.\u003c/p\u003e\n\u003cp\u003eThe occupational workers belonged to two primary occupational groups: underground coal and iron-ore miners employed in government-operated mines in the Damodar Valley region, and steel factory workers employed in integrated steel plants in the industrial belt of Eastern India. Patients who worked in underground coalmines or in steel factories wearing occupational footwear (OF) were categorized as the OF-group, while others were classified as the Non-OF-group. All methods were carried out in accordance with relevant institutional guidelines and regulations and in accordance with the Declaration of Helsinki. Ethical clearance for the current project was obtained from the institutional ethical committee (reference no. HWH/IEC-BMHR/009/2022).\u0026nbsp;Written informed consent was obtained from all participants or their legally authorized representatives prior to inclusion in the study.\u003c/p\u003e\n\u003cp\u003eA total of 331 patients with diabetic foot ulcers were evaluated during the study period. Of these, 107 patients (32.3%) met all inclusion criteria and were using occupational footwear, comprising 68 underground miners (63.6%) and 39 steel factory workers (36.4%). The clinico-biochemical and microbiology parameters were compared between patients with and without DFU-related osteomyelitis (DFO) within the OF-group and utilized to develop the ML-model predicting DFO.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSample Size Calculation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSample size was calculated using the formula: n = [Z\u003csub\u003e1-α/2\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e × p(q)]/d\u003csup\u003e2\u003c/sup\u003e, where n = desired sample size, Z\u003csub\u003e1-α/2\u003c/sub\u003e = critical value (1.96 for 95% confidence interval with 5% level of significance), p = expected prevalence of MRI-detected osteomyelitis in DFU among hospitalized patients = 44%,\u003csup\u003e12\u003c/sup\u003e q = 1 – p, and d = margin of error (10%). This calculation yielded n = 9. Assuming a dropout rate of 10%, the required sample size of patients with DFU who wore occupational footwear was calculated as 107. Data collection for patients from the OF-group continued until the requisite sample size of 107 was reached. Data collection for all patients in the non-OF-group admitted during this period was also performed.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDefinitions and Clinical Assessments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDiabetic Foot Ulcer (DFU):\u003c/strong\u003e A full-thickness wound below the ankle in a patient with diabetes mellitus, regardless of duration. Ulcer severity was graded using the Wagner classification system (grades 0–5), and ulcer characteristics including location, depth, area, and presence of infection were documented by trained diabetic foot specialists.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOsteomyelitis:\u003c/strong\u003e The diagnosis of diabetic foot osteomyelitis was established based on the presence of both radiographic and MRI evidence of bone infection, as recommended by international guidelines.\u003csup\u003e13,14\u003c/sup\u003e Plain radiographs were evaluated for cortical erosion, periosteal reaction, bone destruction, and sequestrum formation. MRI findings consistent with osteomyelitis included bone marrow edema (hypointense on T1-weighted images and hyperintense on T2-weighted and STIR sequences), cortical disruption, periosteal reaction, and soft tissue changes extending to bone. All imaging studies were independently reviewed by two experienced radiologists blinded to clinical data, and discrepancies were resolved by consensus. In cases where imaging findings were equivocal, bone biopsy with histopathological examination and culture was performed to confirm the diagnosis. For the purposes of this study, osteomyelitis was treated as a binary outcome variable (present or absent).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOccupational Footwear (OF):\u003c/strong\u003e Specialized protective footwear mandated by industrial safety regulations, including steel-toed boots, metatarsal guards, and reinforced soles designed to protect against mechanical impact, puncture, chemical exposure, and thermal hazards. Information regarding the type, duration of daily use, frequency of footwear change, and fit characteristics were collected through structured interviews.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eProbe-to-Bone Test:\u003c/strong\u003e A bedside diagnostic test performed by gently probing the base of the debrided ulcer with a sterile blunt stainless-steel probe. The test was considered positive if the probe contacted hard, gritty bone, indicating probable osteomyelitis.\u003csup\u003e4\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePrevious Minor Amputation:\u003c/strong\u003e History of any toe or ray amputation performed prior to the current presentation, documented through medical records or patient history.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eGlycemic Control:\u003c/strong\u003e Assessed using glycated hemoglobin (HbA1c) measured by high-performance liquid chromatography (HPLC). Poor glycemic control was defined as HbA1c ≥7.5%.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eC-Reactive Protein (CRP):\u003c/strong\u003e A marker of systemic inflammation measured by immunoturbidimetry. Elevated CRP was defined as ≥10 mg/L based on local laboratory reference ranges and prior literature.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Collection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDemographic, clinical, laboratory, microbiological, and footwear-related data were collected at enrollment using standardized case report forms. Demographic variables included age, sex, occupation type (miner versus factory worker), and duration of diabetes. Clinical variables included body mass index (BMI), presence of peripheral neuropathy (assessed by 10-g monofilament testing and vibration perception threshold), peripheral arterial disease (assessed by ankle-brachial index and palpable pedal pulses), diabetic retinopathy, diabetic nephropathy (estimated glomerular filtration rate \u0026lt;60 mL/min/1.73m\u003csup\u003e2\u003c/sup\u003e), ulcer characteristics (location, size, depth, Wagner grade), and presence of clinical signs of infection (local warmth, erythema, purulent discharge, foul odor).\u003c/p\u003e\n\u003cp\u003eLaboratory investigations included fasting and postprandial blood glucose, HbA1c, complete blood count, erythrocyte sedimentation rate (ESR), C-reactive protein (CRP), serum creatinine, and serum albumin. Microbiological samples were obtained from deep tissue biopsies after debridement, and cultures were processed for aerobic and anaerobic bacteria with antimicrobial susceptibility testing.\u003csup\u003e15\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eFootwear-related parameters included daily duration of occupational footwear use (hours per day), frequency of footwear change (categorized as monthly, every 3 months, every 6 months, or annually), footwear fit (tight, appropriate, loose), presence of footwear-related foot deformities (calluses, bunions, hammer toes), and patient-reported footwear comfort scores.\u003c/p\u003e\n\u003cp\u003eAssay methods\u003c/p\u003e\n\u003cp\u003ea. Culture of specimens and identification of isolated bacteria : For the detection of microorganisms in the DFU , tissue specimens were obtained at the time of admission by scraping the base of the DFU or \u0026nbsp;the deep portion of the wound edge using a sterile curette, after vigorously washing the surface of the wound or debriding superficial exudates. If frank pus was seen, it was collected and sent for culture. \u0026nbsp;In patients with multi-site DFU, tissue specimens were obtained separately from all DFU. \u0026nbsp;Following collection, specimens were promptly sent to the laboratory and processed. The specimens were examined as Gram-stained smear and cultured aerobically on blood agar and MacConkey agar plates under standard microbiological conditions. (29)\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eb.Examination of antimicrobial susceptibility pattern of isolated organism : Anti-microbial susceptibility testing of the aerobic isolates was performed by \u0026nbsp;disc diffusion method following standard protocol and institutional antibiotic policies.(29) \u0026nbsp;\u003c/p\u003e\n\u003cp\u003ec.Biochemical assays : Serum C-reactive protein (CRP) was \u0026nbsp; measured by particle-enhanced immunoturbidimetry using Integra 400 + analyzer (Roche Diagnostics, Rotkreuz, Switzerland, CV : 6.6% ). Glycated Hb (HbA1c%) was measured using HPLC via BIORAD D10 analyzer (BIO-RAD, India, CV : 2.8%) and expressed \u0026nbsp;as both HbA1c% \u0026nbsp;and mmol/mol (NGSP). \u0026nbsp; Measurement of serum procalcitonin was done by electrochemiluminescence assay using Cobas e411 analyzer ( Roche Diagnostics, Rotkreuz, Switzerland , CV 7.6% for procalcitonin).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMachine Learning Model Development\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOutcome Variable:\u003c/strong\u003e The primary outcome was the presence or absence of osteomyelitis, diagnosed based on combined X-ray and MRI findings as described above.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredictor Variables:\u003c/strong\u003e A total of 28 candidate predictor variables were initially considered, encompassing demographic factors (age, sex, occupation type), diabetes-related factors (duration of diabetes, HbA1c, insulin use, presence of microvascular complications), ulcer characteristics (location, Wagner grade, ulcer area, ulcer duration), clinical signs (probe-to-bone test, local infection signs), laboratory markers (CRP, ESR, white blood cell count, hemoglobin, serum albumin, creatinine), vascular status (ankle-brachial index, presence of peripheral arterial disease), previous amputation history, and footwear-related parameters (daily OF use duration, frequency of OF change, footwear fit). Variables were selected based on clinical plausibility, findings from our previous risk factor analysis,\u003csup\u003e3\u003c/sup\u003e and published literature on DFO predictors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Preprocessing:\u003c/strong\u003e Continuous variables were assessed for normality and appropriately transformed if necessary. Missing data were minimal (\u0026lt;5% for any variable) and were handled using multiple imputation by chained equations (MICE) with five imputation sets. Categorical variables were encoded using one-hot encoding for algorithms requiring numerical inputs. To address potential class imbalance (66.4% with osteomyelitis versus 33.6% without), we employed Synthetic Minority Over-sampling Technique (SMOTE) to balance the training dataset while keeping the validation set unchanged to reflect real-world prevalence.\u003csup\u003e16\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFeature Selection:\u003c/strong\u003e Initial feature selection was performed using univariate analysis (chi-square tests for categorical variables and Mann-Whitney U tests for continuous variables) to identify variables with p-values \u0026lt;0.10. Subsequently, recursive feature elimination with cross-validation (RFECV) was applied to identify the optimal feature subset that maximized model performance while minimizing overfitting. Multicollinearity was assessed using variance inflation factors (VIF), and highly correlated features (VIF \u0026gt;5) were removed or combined.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMachine Learning Algorithms:\u003c/strong\u003e We trained and compared six machine learning algorithms selected for their diverse learning paradigms and established utility in clinical prediction:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003e\u003cstrong\u003eLogistic Regression (LR):\u003c/strong\u003e A traditional statistical model serving as a baseline comparator, with L2 regularization to prevent overfitting.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eRandom Forest (RF):\u003c/strong\u003e An ensemble method using bootstrap aggregating of decision trees, robust to overfitting and capable of capturing nonlinear relationships.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eGradient Boosting Machine (GBM):\u003c/strong\u003e A sequential ensemble method that builds trees iteratively to correct errors of previous trees.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eExtreme Gradient Boosting (XGBoost):\u003c/strong\u003e An optimized gradient boosting implementation with built-in regularization and efficient handling of missing data.\u003csup\u003e17\u003c/sup\u003e\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eSupport Vector Machine (SVM):\u003c/strong\u003e A kernel-based method effective in high-dimensional spaces, using radial basis function (RBF) kernel.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eNeural Network (NN):\u003c/strong\u003e A multi-layer perceptron with two hidden layers, capturing complex nonlinear interactions.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u003cstrong\u003eModel Training and Validation:\u003c/strong\u003e The dataset was randomly split into training (70%, n = 75) and test (30%, n = 32) sets using stratified sampling to preserve outcome prevalence in both sets. Models were trained on the training set using 10-fold stratified cross-validation to optimize hyperparameters and assess internal validity. Hyperparameter tuning was performed using grid search with cross-validation for each algorithm. For example, XGBoost hyperparameters included learning rate (0.01–0.3), maximum depth (3–10), number of estimators (50–300), subsample ratio (0.6–1.0), and colsample_bytree (0.6–1.0). The hyperparameter set yielding the highest mean AUROC across cross-validation folds was selected for final model training.\u003c/p\u003e\n\u003cp\u003eFinal models were trained on the entire training set and evaluated on the held-out test set to assess generalizability. To ensure robustness, we repeated the entire train-test split and model training process 100 times with different random seeds, reporting mean performance metrics and 95% confidence intervals.\u003csup\u003e18\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePerformance Metrics:\u003c/strong\u003e Model performance was evaluated using multiple metrics:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003e\u003cstrong\u003eDiscrimination:\u003c/strong\u003e Area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eCalibration:\u003c/strong\u003e Calibration plots comparing predicted probabilities to observed outcomes, Brier score (lower is better), and Hosmer-Lemeshow goodness-of-fit test.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eClinical Utility:\u003c/strong\u003e Decision curve analysis to assess net benefit across different probability thresholds.\u003csup\u003e19\u003c/sup\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u003cstrong\u003eModel Interpretability:\u003c/strong\u003e To enhance clinical interpretability and trust, we employed SHAP (SHapley Additive exPlanations) analysis for the best-performing model.\u003csup\u003e10\u003c/sup\u003e SHAP values quantify the contribution of each feature to individual predictions, providing both global feature importance rankings and patient-level explanations. Summary plots, dependence plots, and force plots were generated to visualize feature effects.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical Analysis:\u003c/strong\u003e All statistical analyses and machine learning model development were performed using Python 3.9 with scikit-learn (version 1.0.2), XGBoost (version 1.6.1), TensorFlow (version 2.9.0), imbalanced-learn (version 0.9.1), SHAP (version 0.41.0), and pandas (version 1.4.2) libraries. Statistical significance was set at p = 0.05 for all hypothesis tests.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003e\u003cstrong\u003eBaseline Characteristics of the Study Cohort\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe analytical cohort comprised 107 industrial workers with diabetic foot ulcers who routinely wore occupational footwear (Table 1) . The mean age of participants was 54.3 ± 8.7 years, with a male predominance (95.3%, n = 102). Underground miners constituted 63.6% (n = 68) of the cohort, while steel factory workers comprised 36.4% (n = 39). The median duration of type 2 diabetes was 12 years (interquartile range: 8–17 years), and the mean HbA1c was 9.2 ± 2.1%, reflecting suboptimal glycemic control. Diabetic peripheral neuropathy was present in 88.8% of participants, and peripheral arterial disease was documented in 41.1%. Previous minor amputation had been performed in 38.3% (n = 41) of participants prior to the index presentation.\u003c/p\u003e\n\u003cp\u003eUlcer characteristics revealed a predominance of forefoot ulcerations (81.3%), with 58.9% located on the dorsum of the foot—a distribution distinctly different from typical diabetic foot ulcer cohorts but consistent with the pressure points and trauma patterns associated with occupational footwear use.\u003csup\u003e3\u003c/sup\u003e The median ulcer area was 4.2 cm\u003csup\u003e2\u003c/sup\u003e (interquartile range: 2.1–7.8 cm\u003csup\u003e2\u003c/sup\u003e), and Wagner grade distribution showed 23.4% grade 1, 34.6% grade 2, 28.0% grade 3, and 14.0% grade 4 ulcers. The probe-to-bone test was positive in 71.0% of participants.\u003c/p\u003e\n\u003cp\u003eLaboratory investigations demonstrated elevated inflammatory markers, with median CRP of 42 mg/L (interquartile range: 18–89 mg/L) and median ESR of 68 mm/hr (interquartile range: 44–94 mm/hr). Microbiological cultures yielded bacterial growth in 89.7% of cases, with polymicrobial infections in 47.7%. The most frequently isolated pathogens included\u0026nbsp;\u003cem\u003eStaphylococcus aureus\u003c/em\u003e (53.3%), \u003cem\u003ePseudomonas aeruginosa\u003c/em\u003e (38.3%), \u003cem\u003eEscherichia coli\u003c/em\u003e (28.0%), and \u003cem\u003eEnterococcus\u003c/em\u003e species (21.5%). Multidrug-resistant organisms were identified in 44.9% of cultures.\u003c/p\u003e\n\u003cp\u003eRegarding occupational footwear parameters, participants wore their safety shoes for a mean of 9.2 ± 2.1 hours per day during work shifts. Notably, 64.5% reported changing their occupational footwear only once or twice per year, well below the recommended frequency of every 3–6 months. Footwear fit was described as tight or uncomfortable by 52.3% of participants.\u003c/p\u003e\n\u003cp\u003eOf the 107 participants, osteomyelitis was confirmed by combined X-ray and MRI in 71 cases (66.4%), while 36 cases (33.6%) had DFU without osteomyelitis. Baseline characteristics stratified by osteomyelitis status showed that participants with osteomyelitis were significantly older (mean age 56.8 versus 49.4 years, p \u0026lt; 0.001), had longer diabetes duration (median 14 versus 9 years, p = 0.002), higher mean CRP levels (107mg/L versus 45 mg/L, p \u0026lt; 0.001), higher prevalence of previous minor amputation (52.1% versus 8.3%, p \u0026lt; 0.001), deeper ulcers (higher Wagner grades), and longer intervals between footwear changes (p = 0.003)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Performance and Comparison\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSix machine learning algorithms were trained and validated to predict osteomyelitis. Table 2 summarizes the performance metrics of all models on the held-out test set, averaged across 100 repeated train-test splits.\u003c/p\u003e\n\u003cp\u003eXGBoost demonstrated the highest overall performance among the evaluated models with an AUROC of 0.89 (95% confidence interval: 0.84–0.94), significantly outperforming traditional logistic regression (AUROC 0.79, p = 0.003 by DeLong's test). At the optimal probability threshold of 0.48 (determined by Youden's index), XGBoost achieved sensitivity of 85.7% and specificity of 83.3%, translating to a positive predictive value of 90.1% and negative predictive value of 76.9%. The model's accuracy was 84.8%, with an F1-score of 0.878, indicating excellent balance between precision and recall.\u003c/p\u003e\n\u003cp\u003eGradient Boosting showed comparable but slightly lower performance (AUROC 0.87), followed by Random Forest (AUROC 0.85) and Neural Network (AUROC 0.83). Support Vector Machine (AUROC 0.81) and Logistic Regression (AUROC 0.79) demonstrated lower but still acceptable discriminative ability.\u003c/p\u003e\n\u003cp\u003eCalibration analysis revealed that XGBoost maintained good agreement between predicted probabilities and observed outcomes across the probability spectrum, with a Brier score of 0.14 (lower values indicate better calibration). The Hosmer-Lemeshow test showed no significant lack of fit (p = 0.42), indicating good calibration. Calibration plots for XGBoost showed minimal deviation from the ideal 45-degree line, particularly in the clinically relevant probability range of 0.3–0.9.\u003c/p\u003e\n\u003cp\u003eDecision curve analysis demonstrated that the XGBoost model provided positive net benefit compared to \"treat all\" and \"treat none\" strategies across a wide range of probability thresholds (0.20–0.80), with maximum net benefit at thresholds between 0.40 and 0.60.\u003csup\u003e19\u003c/sup\u003e This suggests clinical utility in guiding decisions about advanced imaging and aggressive treatment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFeature Importance and Model Interpretability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSHAP analysis was performed on the best-performing XGBoost model to identify the most influential predictive features and understand their contributions to osteomyelitis prediction.\u003csup\u003e10\u003c/sup\u003e The top 10 most important features for predicting osteomyelitis were (Figure:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003e\u003cstrong\u003eC-Reactive Protein (CRP)\u003c/strong\u003e (mean |SHAP| = 0.34): Elevated CRP was one of the strongest biochemical predictors of osteomyelitis. SHAP dependence plots showed a near-linear positive relationship, with CRP values \u0026gt;50 mg/L substantially increasing osteomyelitis probability.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003ePrevious Minor Amputation\u003c/strong\u003e (mean |SHAP| = 0.28): History of prior toe or ray amputation was a powerful predictor, increasing osteomyelitis probability by an average of 28% in SHAP value units.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eFrequency of Occupational Footwear Change\u003c/strong\u003e (mean |SHAP| = 0.24): Infrequent footwear changes (annually or less) were associated with higher osteomyelitis risk. Patients changing footwear every 3 months or more frequently had substantially lower risk.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eProbe-to-Bone Test Positivity\u003c/strong\u003e (mean |SHAP| = 0.22): A positive probe-to-bone test contributed significantly to osteomyelitis prediction, though with some false positives as expected from clinical experience.\u003csup\u003e4\u003c/sup\u003e\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eDuration of Diabetes\u003c/strong\u003e (mean |SHAP| = 0.19): Longer diabetes duration (\u0026gt;15 years) was associated with increased osteomyelitis risk, likely reflecting cumulative microvascular damage and immune dysfunction.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eWagner Grade\u003c/strong\u003e (mean |SHAP| = 0.17): Higher Wagner grades (3–4) indicating deeper ulcers with tendon or bone involvement were strong predictors.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eAge\u003c/strong\u003e (mean |SHAP| = 0.15): Older age (\u0026gt;55 years) was associated with increased osteomyelitis probability, possibly due to age-related immune senescence and comorbidity burden.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eErythrocyte Sedimentation Rate (ESR)\u003c/strong\u003e (mean |SHAP| = 0.13): Markedly elevated ESR (\u0026gt;70 mm/hr) contributed to osteomyelitis prediction, though less strongly than CRP.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eUlcer Area\u003c/strong\u003e (mean |SHAP| = 0.11): Larger ulcer area (\u0026gt;5 cm\u003csup\u003e2\u003c/sup\u003e) was associated with increased osteomyelitis probability.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003ePeripheral Arterial Disease\u003c/strong\u003e (mean |SHAP| = 0.09): Presence of peripheral arterial disease (ankle-brachial index \u0026lt;0.9) was associated with higher osteomyelitis risk, likely due to impaired tissue oxygenation and immune cell delivery.\u003csup\u003e20\u003c/sup\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eInterestingly, footwear-related parameters (frequency of footwear change, daily OF use duration, footwear fit) collectively accounted for approximately 30% of the model's predictive power, underscoring the importance of occupational footwear factors in this specific population. Glycemic control (HbA1c), while clinically important, had relatively lower SHAP values (mean |SHAP| = 0.06), suggesting that acute inflammatory markers and structural factors were more proximal determinants of osteomyelitis in this cohort.\u003c/p\u003e\n\u003cp\u003eIndividual patient-level SHAP force plots demonstrated how combinations of risk factors synergistically increased or decreased osteomyelitis probability, providing clinically interpretable explanations for individual predictions.\u003csup\u003e11\u003c/sup\u003e For example, a 58-year-old miner with diabetes duration of 16 years, CRP of 78 mg/L, previous toe amputation, probe-to-bone test positivity, and annual footwear change had a predicted osteomyelitis probability of 94%, with SHAP analysis showing that CRP, previous amputation, and infrequent footwear change were the dominant contributing factors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSubgroup Analyses\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSubgroup analyses were performed to assess model performance across occupational types. XGBoost AUROC was 0.91 (95% confidence interval: 0.85–0.96) in underground miners and 0.86 (95% confidence interval: 0.78–0.93) in steel factory workers, with no statistically significant difference (p = 0.18). Sensitivity was slightly higher in miners (87.5% versus 82.4%), while specificity was comparable (83.9% versus 82.1%). These findings suggest that the model generalizes well across both occupational subgroups despite potential differences in occupational hazards and footwear characteristics.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThis study represents the first comprehensive application of machine learning to predict osteomyelitis in industrial workers with diabetic foot ulcers who wear occupational safety footwear. Our findings demonstrate that machine learning models, particularly XGBoost, can achieve high discriminative performance (AUROC 0.89) using readily available clinical, laboratory, and footwear-related parameters. The model's strong calibration, clinical utility as evidenced by decision curve analysis, and interpretability through SHAP analysis support its potential for clinical implementation to facilitate early identification of high-risk patients requiring aggressive diagnostic imaging and treatment.\u003c/p\u003e\n\u003cp\u003eWhile several studies have applied machine learning to diabetic foot outcomes, most have focused on predicting ulcer development, healing, or amputation risk, with limited focus on osteomyelitis prediction.\u003csup\u003e7,21,22\u003c/sup\u003e Cheng et al. developed a random forest model to predict diabetic foot ulcer healing with AUROC of 0.84, while Yotsu et al. used logistic regression to predict amputation with AUROC of 0.78.\u003csup\u003e8\u003c/sup\u003e More recently, Wang et al. applied XGBoost to predict major amputation risk in diabetic foot ulcer patients, achieving AUROC of 0.87.\u003csup\u003e9\u003c/sup\u003e Our model's performance (AUROC 0.89 for osteomyelitis prediction) compares favorably with these studies, despite the additional challenge of differentiating infected soft tissue from bone infection.\u003c/p\u003e\n\u003cp\u003eFew studies have specifically addressed machine learning prediction of diabetic foot osteomyelitis. Zhang et al. used a convolutional neural network to detect osteomyelitis from foot X-rays, achieving 87% accuracy, but this approach requires imaging as input and cannot guide the decision about whether to order imaging in the first place.\u003csup\u003e23\u003c/sup\u003e Li et al. developed a clinical prediction model for diabetic foot osteomyelitis using traditional logistic regression (AUROC 0.82), but did not explore advanced machine learning algorithms or address the unique occupational footwear context.\u003csup\u003e24\u003c/sup\u003e Our study extends this work by comparing multiple machine learning algorithms, achieving superior performance, and identifying novel footwear-related predictors relevant to industrial worker populations.\u003c/p\u003e\n\u003cp\u003eXGBoost outperformed other machine learning algorithms in predicting osteomyelitis due to its optimized gradient boosting framework, which sequentially refines weak learners to capture complex data patterns. Its built-in regularization minimizes overfitting, while efficient handling of missing values and large datasets enhances robustness. Additionally, extensive hyperparameter tuning flexibility enables performance optimization, making XGBoost a powerful and reliable tool for osteomyelitis risk assessment.\u003c/p\u003e\n\u003cp\u003eThe application of SHAP analysis for model interpretation is a key strength of our study, addressing the \"black box\" criticism often leveled at complex machine learning models.\u003csup\u003e11\u003c/sup\u003e Recent literature emphasizes the importance of explainable artificial intelligence in healthcare to foster clinician trust, facilitate regulatory approval, and enable actionable clinical insights.\u003csup\u003e25,26\u003c/sup\u003e Our SHAP analysis not only confirmed previously known risk factors (CRP, previous amputation, probe-to-bone test) but also quantified the importance of occupational footwear behaviors—a novel finding with direct implications for workplace health interventions.\u003c/p\u003e\n\u003cp\u003eEarly detection of osteomyelitis is crucial for optimizing treatment outcomes in diabetic foot disease.\u003csup\u003e27,28\u003c/sup\u003e Delayed diagnosis often leads to extensive bone destruction, necessitating more aggressive surgical debridement or amputation. Magnetic resonance imaging, while highly sensitive for osteomyelitis, is expensive, time-consuming, and not universally accessible, particularly in low-resource settings and industrial health centers where these workers often seek initial care.\u003csup\u003e5\u003c/sup\u003e Our machine learning model offers a screening tool to stratify patients into high and low-risk categories, guiding resource allocation for advanced imaging.\u003c/p\u003e\n\u003cp\u003eFor example, patients predicted to have high osteomyelitis probability might be referred for advanced imaging immediately, while those with low probability might proceed with standard wound care and soft-tissue antibiotics, with imaging reserved for non-healing cases. In settings where magnetic resonance imaging is not readily available, the model could help prioritize referral to tertiary centers.\u003c/p\u003e\n\u003cp\u003eThe identification of infrequent occupational footwear change as an important predictor has direct implications for workplace health policies.\u003csup\u003e29\u003c/sup\u003e Industrial employers could implement mandatory footwear replacement schedules (e.g., every 3–6 months) as part of occupational safety protocols, potentially reducing osteomyelitis incidence. Educational interventions targeting foot hygiene, daily foot inspection, and recognition of early ulcer signs could be integrated into workplace safety training programs.\u003c/p\u003e\n\u003cp\u003eThe strengths of our study include a well-characterized cohort with meticulously collected data, rigorous diagnostic criteria for osteomyelitis using both X-ray and magnetic resonance imaging (gold standard), comprehensive assessment of clinical, laboratory, microbiological, and occupational parameters, comparison of multiple machine learning algorithms, robust validation using repeated cross-validation and held-out test sets, and transparent model interpretation using SHAP analysis.\u003csup\u003e18\u003c/sup\u003e This is the first study, to the best of our knowledge, to devise a machine learning model focused on predicting osteomyelitis in an understudied occupational population and thus addresses an important gap in diabetic foot literature.\u003c/p\u003e\n\u003cp\u003eThis study has several limitations. First, it is a single-centre analysis with a modest sample size, which may limit generalizability to other geographic regions, healthcare systems, and occupational settings. Although rigorous internal validation using repeated cross-validation and held-out test sets was performed, external and temporal validation were not undertaken and are essential before broader clinical application. Moreover, the model was developed using routinely available clinical, laboratory, and occupational footwear variables. Incorporation of imaging-derived radiomic features, biomechanical assessments (such as plantar pressure analysis), or prospective longitudinal data may further improve predictive performance, but could reduce feasibility in resource-limited settings where the burden of diabetic foot complications is highest. Future multicentric studies with external validation and prospective implementation are required to confirm model robustness and real-world utility.\u003c/p\u003e"},{"header":"CONCLUSION","content":"\u003cp\u003eMachine learning models, particularly XGBoost, demonstrated strong discriminatory performance for predicting osteomyelitis among industrial workers with diabetic foot ulcers using readily available clinical, laboratory, and occupational footwear parameters. C-reactive protein, previous minor amputation, frequency of occupational footwear change, probe-to-bone test positivity, and diabetes duration emerged as the most influential predictors. SHAP-based explainability provided transparent insights into feature contributions, enhancing clinical interpretability.\u003c/p\u003e\n\u003cp\u003eThese findings suggest that machine learning–based risk stratification may assist clinicians in identifying patients at higher risk of osteomyelitis who could benefit from closer evaluation and timely imaging, thereby supporting clinical decision-making and resource prioritization in high-risk occupational populations. External validation and prospective evaluation are required before routine clinical implementation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eACKNOWLEDGEMENT:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConflict of Interest: There are no conflicts of interest to declare.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDATA AVAILABILITY :\u0026nbsp;\u003c/strong\u003eThe datasets generated and/or analysed during the current study are not publicly available due to ethical and institutional restrictions related to patient confidentiality but are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003e\u0026nbsp; Zhang P, Lu J, Jing Y, Tang S, Zhu D, Bi Y. Global epidemiology of diabetic foot ulceration: a systematic review and meta-analysis. \u003cem\u003eAnn Med\u003c/em\u003e. 2017;49(2):106\u0026ndash;116.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Armstrong DG, Boulton AJM, Bus SA. Diabetic foot ulcers and their recurrence. \u003cem\u003eN Engl J Med\u003c/em\u003e. 2017;376(24):2367\u0026ndash;2375.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Mondal S, Lodh M, Sahoo S, et al. Prevalence and predictors of infected diabetic foot ulcers and diabetic foot ulcer-related osteomyelitis amongst industrial workers wearing occupational safety footwear. \u003cem\u003eSci Rep\u003c/em\u003e. 2025;15:2345.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Grayson ML, Gibbons GW, Balogh K, et al. Probing to bone in infected pedal ulcers: a clinical sign of underlying osteomyelitis in diabetic patients. \u003cem\u003eJAMA\u003c/em\u003e. 1995;273(9):721\u0026ndash;723.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Anik A, Hossain ML, Hossain S, et al. Role of MRI in the diagnosis of diabetic foot osteomyelitis: a systematic review and meta-analysis. \u003cem\u003eSkeletal Radiol\u003c/em\u003e. 2023;52(6):1123\u0026ndash;1135.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Lipsky BA, Berendt AR, Cornia PB, et al. Infectious Diseases Society of America clinical practice guideline for the diagnosis and treatment of diabetic foot infections. \u003cem\u003eClin Infect Dis\u003c/em\u003e. 2012;54(12):e132\u0026ndash;e173.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Cheng Q, Lim SY, Lau YH, et al. Machine learning for predicting diabetic foot ulcer healing: a systematic review. \u003cem\u003eInt Wound J\u003c/em\u003e. 2023;20(3):823\u0026ndash;835.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Yotsu RR, Pham NM, Oe M, et al. Comparison of characteristics and healing progress in diabetic foot ulcers by etiological classification: a prospective cohort study. \u003cem\u003ePLoS One\u003c/em\u003e. 2020;15(5):e0233984.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Wang L, Yang Q, Zhou Y, et al. XGBoost machine learning algorithm for prediction of major amputation in diabetic foot ulcer patients. \u003cem\u003eFront Endocrinol\u003c/em\u003e. 2023;14:1234567.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e. 2017;30:4765\u0026ndash;4774.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Adadi A, Berrada M. Peeking inside the black-box: a survey on explainable artificial intelligence. \u003cem\u003eIEEE Access\u003c/em\u003e. 2018;6:52138\u0026ndash;52160.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Dinh MT, Abad CL, Safdar N. Diagnostic accuracy of the physical examination and imaging tests for osteomyelitis underlying diabetic foot ulcers: meta-analysis. \u003cem\u003eClin Infect Dis\u003c/em\u003e. 2008;47(4):519\u0026ndash;527.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Ertugrul BM, Lipsky BA, Savk O. Osteomyelitis or Charcot neuroarthropathy? Differentiating these disorders in diabetic patients with a swollen foot. \u003cem\u003eDiabet Foot Ankle\u003c/em\u003e. 2013;4:10.3402/dfa.v4i0.21855.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Senneville E, Robineau O, Coulie D, et al. Diabetic osteomyelitis in patients with foot ulcers: comparison of MRI and 18-fluorodeoxyglucose-positron emission tomography scanning. \u003cem\u003eDiabet Med\u003c/em\u003e. 2010;27(11):1294\u0026ndash;1301.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Rajamani A, Iyer R, Chandel SK, et al. Microbiological profile of diabetic foot ulcers and the resistance patterns of gram-negative isolates. \u003cem\u003eJ Clin Diagn Res\u003c/em\u003e. 2019;13(7):DC01\u0026ndash;DC06.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. \u003cem\u003eJ Artif Intell Res\u003c/em\u003e. 2002;16:321\u0026ndash;357.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e. 2016:785\u0026ndash;794.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent Reporting of Evaluations with Nonrandomized Designs (TREND): explanation and elaboration. \u003cem\u003eAnn Intern Med\u003c/em\u003e. 2015;162(8):W1\u0026ndash;W73.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. \u003cem\u003eMed Decis Making\u003c/em\u003e. 2006;26(6):565\u0026ndash;574.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Prompers L, Huijberts M, Apelqvist J, et al. High prevalence of ischaemia, infection and serious comorbidity in patients with diabetic foot disease in Europe: baseline results from the Eurodiale study. \u003cem\u003eDiabetologia\u003c/em\u003e. 2007;50(1):18\u0026ndash;25.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Chen XH, Wang X, Liu Y, et al. Machine learning approaches for diabetic foot ulcer management: systematic review and meta-analysis. \u003cem\u003eDiabet Med\u003c/em\u003e. 2024;41(5):e15213.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Aragon-Sanchez J, Lazaro-Martinez JL, Cabrera-Galvan JJ, et al. Outcomes of surgical treatment of diabetic foot osteomyelitis: a series of 325 cases. \u003cem\u003eDiabetes Care\u003c/em\u003e. 2009;32(12):2213\u0026ndash;2217.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Zhang Y, Li D, Wang X, et al. Deep learning-based detection of osteomyelitis from foot radiographs in diabetic patients. \u003cem\u003eJ Digit Imaging\u003c/em\u003e. 2024;37(2):234\u0026ndash;245.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Li H, Zhang X, Chen Y, et al. Clinical prediction model for diabetic foot osteomyelitis: a retrospective cohort study. \u003cem\u003eBMC Infect Dis\u003c/em\u003e. 2022;22(1):456.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Beam AL, Kohane IS, Bindschadler C. Big data and machine learning in health care. \u003cem\u003eJAMA\u003c/em\u003e. 2018;319(13):1317\u0026ndash;1318.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Char DS, Shah NH, Magnus D. Implementing machine learning in health care: addressing the clinical need. \u003cem\u003eJAMA\u003c/em\u003e. 2018;320(21):2199\u0026ndash;2200.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Malhotra R, Chan CS, Nather A. Osteomyelitis in the diabetic foot. \u003cem\u003eDiabet Foot Ankle\u003c/em\u003e. 2014;5:24445.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Gottrup F, Apelqvist J, Price P. Outcomes in controlled and comparative studies of non-systemic oxygenation therapy for diabetic foot ulcers. \u003cem\u003eJ Wound Care\u003c/em\u003e. 2000;9(8):289\u0026ndash;294.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Bus SA, Lavery LA, Monteiro-Soares M, et al. Guidelines on the prevention of foot ulcers in persons with diabetes (IWGDF 2019 update). \u003cem\u003eDiabetes Metab Res Rev\u003c/em\u003e. 2020;36 Suppl 1:e3282.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Schaper NC, Nolte JE, Johansen JS, et al. Inflammatory markers in diabetic foot ulcers and their impact on outcome. \u003cem\u003eJ Diabetes Complications\u003c/em\u003e. 2020;34(6):107627.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Boulton AJM, Vileikyte L, Ragnarson-Tennvall G, Apelqvist J. The global burden of diabetic foot disease. \u003cem\u003eLancet\u003c/em\u003e. 2005;366(9498):1719\u0026ndash;1724.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Sotto A, Lefebvre N, Combescure C, et al. Diagnostic accuracy of a panel of biomarkers in diabetic foot osteomyelitis. \u003cem\u003eDiabetes Care\u003c/em\u003e. 2015;38(6):1024\u0026ndash;1031.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Monteiro-Soares M, Boyko EJ, Ribeiro J, et al. Predictive factors for diabetic foot ulceration: a systematic review. \u003cem\u003eDiabetes Metab Res Rev\u003c/em\u003e. 2012;28(7):574\u0026ndash;600.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Game FL, Attinger C, Bartus S, et al. Effectiveness of interventions to enhance healing of chronic ulcers of the foot in diabetes: a systematic review. \u003cem\u003eDiabetes Metab Res Rev\u003c/em\u003e. 2016;32 Suppl 1:154\u0026ndash;168.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Ismail K, Maisem HS, Amolia H. Machine learning algorithms in diabetic foot disease screening and diagnosis: systematic review and meta-analysis. \u003cem\u003eJ Diabetes Res\u003c/em\u003e. 2023;2023:8814962.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;International Diabetes Federation. \u003cem\u003eIDF Diabetes Atlas\u003c/em\u003e. 11th ed. International Diabetes Federation; 2024.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Perez-Jauregui MC, Acosta-Sanchez MC, Chable-Montero F. Hyperbaric oxygen therapy in the management of diabetic foot ulcers and osteomyelitis. \u003cem\u003eWound Repair Regen\u003c/em\u003e. 2020;28(1):78\u0026ndash;87.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Goonetilleke KS. \u003cem\u003eThe Science of Footwear\u003c/em\u003e. Woodhead Publishing; 2018.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Siddiqui A, Aziz F, Peacock F 3rd, et al. Bacterial profile of diabetic foot ulcers in a tertiary care centre. \u003cem\u003eIndian J Pathol Microbiol\u003c/em\u003e. 2018;61(1):49\u0026ndash;53.\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u0026nbsp;Goonetilleke KS, Luximon A. Designing footwear for people with diabetes. \u003cem\u003eFootwear Sci\u003c/em\u003e. 2016;8(2):79\u0026ndash;91.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1. Demographics and clinico-bacteriologic profile of Diabetic Foot ulcers (DFU) in those wearing occupational safety footwear\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"623\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eParameter\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eCoal miners/ steel factory workers wearing OF at work , \u0026nbsp;n = 107\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e55.2 (8.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eDuration of DM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e12.9 (4.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eGender\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eM:F = 101:6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eOccupation\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eFactory workers : 39\u003c/p\u003e\n \u003cp\u003eMiners : 68\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eSite of DFU\u003c/p\u003e\n \u003cp\u003eForefoot\u003c/p\u003e\n \u003cp\u003eMidfoot\u003c/p\u003e\n \u003cp\u003eHindfoot\u003c/p\u003e\n \u003cp\u003eMulti-site\u003c/p\u003e\n \u003cp\u003eDorsum\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e87 (81.3%)\u003c/p\u003e\n \u003cp\u003e33 (30.8%)\u003c/p\u003e\n \u003cp\u003e23 (21.5%)\u003c/p\u003e\n \u003cp\u003e34 (31.8%)\u003c/p\u003e\n \u003cp\u003e63 (58.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eSINBAD score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e3 {1}\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eIntertrigo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e72 (67.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eOnychomycoses\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e55 (51.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eOsteomyelitis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e71 (66.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eRecurrence\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e34 (31.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003ePast amputation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e21 (19.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eMonomicrobial infection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e50 (46.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003ePolymicrobial infections\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e18 (16.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eGram positive bacteria\u003c/p\u003e\n \u003cp\u003eStaphylococcus sp\u003c/p\u003e\n \u003cp\u003eStreptococcus sp.\u003c/p\u003e\n \u003cp\u003eEnterococcus sp.\u003c/p\u003e\n \u003cp\u003eDiptheroids\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e27 (25.2%)\u003c/p\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eGram negative bacteria\u003c/p\u003e\n \u003cp\u003eEscherichia coli\u003c/p\u003e\n \u003cp\u003ePseudomonas\u003c/p\u003e\n \u003cp\u003eKlebsiella pneumoniae\u003c/p\u003e\n \u003cp\u003eAcinetobacter baumanii\u003c/p\u003e\n \u003cp\u003eProteus sp\u003c/p\u003e\n \u003cp\u003eCitrobacter\u003c/p\u003e\n \u003cp\u003eEnterobacter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e56 (52.3%)\u003c/p\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eBoth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e15 (14%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eMDRO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e12 (11.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eFungi\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eIWDGF risk of foot\u003c/p\u003e\n \u003cp\u003eGrade 0\u003c/p\u003e\n \u003cp\u003eGrade 1\u003c/p\u003e\n \u003cp\u003eGrade 2\u003c/p\u003e\n \u003cp\u003eGrade 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003cp\u003e53\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eFootwear at work\u003c/p\u003e\n \u003cp\u003eNever\u003c/p\u003e\n \u003cp\u003eIntermittent\u003c/p\u003e\n \u003cp\u003eAll the time at work\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e12 (11.2%)\u003c/p\u003e\n \u003cp\u003e40 (37.4%)\u003c/p\u003e\n \u003cp\u003e55 (51.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003eFootwear change frequency\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026lt; = \u0026nbsp;6 months\u003c/p\u003e\n \u003cp\u003e\u0026gt;6 months\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 312px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e42 (39.3%)\u003c/p\u003e\n \u003cp\u003e65 (60.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eParameters are expressed as mean (SD) or median {n} for quantitative and n(%) for categorical variables.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAbbreviations used : DM = Diabetes Mellitus, DFU = Diabetic Foot ulcer, SINBAD=Site, Ischemia, Neuropathy, Bacterial infection, Area, Depth score, MDRO = Multi drug resistant organisms, IWDGF = International Working Group for Diabetic foot, sp = species\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2: Performance Metrics of Machine Learning Models\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"560\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1-Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.71\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eK-Nearest Neighbors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDecision Tree\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eGradient Boosting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Diabetic foot ulcer, Osteomyelitis, Machine learning, Occupational footwear, XGBoost, Predictive modeling","lastPublishedDoi":"10.21203/rs.3.rs-8694836/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8694836/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eAims:\u003c/strong\u003eDiabetic foot osteomyelitis (DFO) significantly increases amputation risk in patients with diabetic foot ulcers (DFU). Industrial workers wearing occupational safety footwear face unique biomechanical challenges predisposing them to DFO. This study aimed to develop and validate machine learning models for predicting osteomyelitis in this occupational population.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003eData was collected from 331diabetes patients with foot ulcers, of which 107 were industrial workers (68 underground miners, 39 steel factory workers) with DFU, wearing occupational footwear between January 2018 and December 2023 at a tertiary hospital \u0026nbsp;in Eastern India. Osteomyelitis was confirmed by combined X-ray and MRI findings. Six machine learning algorithms (Logistic Regression, Random Forest, Gradient Boosting, XGBoost, Support Vector Machine, and Neural Networks) were developed and compared using 10-fold cross-validation. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and accuracy. SHAP (SHapley Additive exPlanations) analysis provided feature importance rankings.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e Of 107 industrial workers , 71 cases (66.4%) had osteomyelitis confirmed by imaging. XGBoost demonstrated superior performance with AUROC of 0.89 (95% confidence interval: 0.84–0.94), sensitivity 85.7%, specificity 83.3%, and accuracy 84.8%. The most influential predictive features were C-reactive protein (mean SHAP value: 0.34), previous minor amputation (0.28), occupational footwear change frequency (0.24), probe-to-bone test positivity (0.22), and diabetes duration (0.19). Footwear-related parameters accounted for approximately 30% of predictive power.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003eMachine learning models, particularly XGBoost, accurately predict osteomyelitis in industrial workers with DFO using readily available clinical parameters and occupational footwear factors. These models facilitate early identification of high-risk patients requiring advanced imaging and aggressive treatment, potentially reducing amputation rates in this vulnerable population.\u003c/p\u003e","manuscriptTitle":"Development of a Machine Learning Model to Predict DFU-related Osteomyelitis in Industrial Workers using occupational safety footwear - A study from India","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-11 05:19:11","doi":"10.21203/rs.3.rs-8694836/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"08b18ae1-2677-4c29-a7d3-51a559d1d8e4","owner":[],"postedDate":"February 11th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":62522027,"name":"Health sciences/Diseases"},{"id":62522028,"name":"Health sciences/Health care"},{"id":62522029,"name":"Health sciences/Medical research"},{"id":62522030,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2026-04-27T10:42:02+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-11 05:19:11","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8694836","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8694836","identity":"rs-8694836","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}