Explainable Artificial Intelligence System Predicts Multiple Diabetic Foot Ulcer Outcomes

A multicentre derivation and validation study accompanied by a prospective randomised clinical evaluation has demonstrated that an explainable artificial intelligence ensemble can simultaneously forecast four major prognostic outcomes in patients hospitalised with diabetic foot ulcers. Reporting in eClinicalMedicine, researchers developed and externally validated a dual-branch machine learning architecture designed to jointly estimate individualised probabilities for wound cure, ulcer recurrence, lower-extremity amputation, and all-cause mortality during observed follow-up. The computational framework integrated retrospective electronic medical record information extracted within the first twenty-four hours of hospital admission across distinct medical centres, combining parallel tree-based and neural network algorithms to overcome the limitations of conventional single-outcome prognostic tools. To assess real-world utility beyond static retrospective datasets, the investigators subsequently embedded the locked algorithmic framework into a web-based clinical platform and conducted a prospective randomised comparative study across three primary care hospitals, evaluating how algorithmic guidance influenced clinician recommendation quality and short-term patient engagement metrics.

Diabetic foot ulcers represent one of the most debilitating microvascular and macrovascular complications arising from chronic hyperglycemia, imposing an immense clinical and economic burden on limb preservation and patient survival. Epidemiological data indicate that diabetic foot ulcers affect an estimated 18.6 million individuals with diabetes worldwide each year, with approximately 19 percent to 34 percent of diabetic individuals developing an ulcer over their lifetime. Following diagnosis, between 50 percent and 65 percent of patients experience ulcer recurrence within three to five years, nearly 20 percent undergo lower-extremity amputation, and one-year mortality rates exceed 10 percent. Standard medical management relies on structured glycemic control, aggressive local wound care, early foot screening, infection control, and pressure offloading. However, because ulcer progression is governed by intricate interactions among systemic metabolic disturbances, peripheral neuropathy, peripheral arterial ischemia, infection, and patient treatment adherence, single clinical indicators frequently fail to capture comprehensive prognostic risk across interdependent clinical endpoints.

Although machine learning algorithms have emerged as powerful tools capable of modeling nonlinear associations and high-dimensional feature interactions in heterogeneous clinical data, earlier prognostic models in diabetic foot care suffered from notable design limitations. A systematic literature search conducted from database inception to April 2026 revealed that existing predictive studies predominantly focused on single clinical endpoints in isolation, such as wound healing, amputation, or in-hospital death, while largely ignoring post-discharge recurrence or multi-outcome interdependencies. Furthermore, published algorithms were frequently constrained by single-centre patient cohorts, modest sample sizes, absent or inadequate external validation, opaque black-box formulations lacking clinical interpretability, and a lack of prospective evaluation within active clinical workflows. Prior models also neglected shared prognostic factors that link simultaneous or sequential complications, thereby missing clinical information across distinct phases of the disease trajectory.

To overcome these structural challenges, the investigators designed a multicentre retrospective investigation drawing upon electronic medical records and longitudinal follow-up data from 863 unique adult patients treated across three anonymised hospitals in China between January 2018 and April 2025. Patients admitted with documented diabetic foot ulcers were identified across Centre A in Jiangsu Province (458 patients), Centre B in Anhui Province (305 patients), and Centre C in Shandong Province (100 patients). After linking repeat admissions to preserve exactly one consolidated record per patient, baseline candidate variables encompassing demographics, laboratory test results, behavioural factors, medical history, and initial treatment characteristics were extracted from the first twenty-four hours of the index admission. Outcomes were tracked through longitudinal hospital records and structured telephone follow-up through September 2025, yielding a median retrospective follow-up duration of 41.2 months with an interquartile range of 20.5 to 60.0 months. Race and ethnicity data were not routinely captured across the participating centres and were therefore not modeled.

To evaluate model generalisability rigorously under geographical transportability constraints, data from Centre A and Centre B comprising 763 patients were pooled exclusively for model development, data preprocessing, and internal cross-validation, whereas the 100 patients from Centre C were completely quarantined as an unseen external validation cohort. Addressing the pervasive clinical reality of irregular examinations and incomplete documentation, the researchers implemented a structured missingness and data-augmentation pipeline. Instead of relying purely on conventional statistical imputation, numerical features were expanded into augmented pairs comprising an observability indicator mask and an imputed value sampled from a uniform distribution, while categorical variables underwent one-hot encoding accompanied by selective group-level masking. Experimental evaluations confirmed that avoiding indiscriminate masking of one-hot encoded categorical blocks preserved the underlying mutually exclusive categorical structure, creating an augmented training distribution that bolstered algorithmic robustness against incomplete clinical profiles without degrading predictive fidelity.

The resulting prediction architecture established a parallel, dual-branch ensemble combining single-task extreme gradient boosting (ST-XGBoost) and single-task multilayer perceptrons (ST-MLP) across each of the four target outcomes. Rather than forcing hard parameter sharing within a joint multi-task neural network—which frequently triggers negative task interference during optimisation—the parallel formulation allowed each clinical endpoint to learn independent feature weights while processing a standardised input space. Within the gradient-boosted decision tree branch, a shared unsupervised autoencoder generated low-dimensional latent embeddings that were concatenated directly with original baseline features prior to model training. Final event probabilities were computed through task-specific linear convex combinations of the two branch outputs, governed by validation-tuned weighting coefficients: 0.11 for wound cure, 0.62 for ulcer recurrence, 0.83 for lower-extremity amputation, and 0.39 for all-cause mortality. This linear late-fusion strategy mitigated overfitting risks on the moderate development sample while safeguarding the transparent tractability of individual branch predictions.

When compared against six alternative modeling approaches—including standard logistic regression, support vector machines, random forests, standalone ST-XGBoost, standalone ST-MLP, and multi-task multilayer perceptrons—the locked ensemble achieved the highest overall multidimensional ranking score of 55 in independent external validation at Centre C. The ensemble model attained a macro-averaged external area under the receiver operating characteristic curve of 0.901, an overall accuracy of 0.875, and a macro-averaged F1 score of 0.830. Task-specific discrimination in the external validation cohort was highest for lower-extremity amputation, yielding an area under the curve of 0.9531 with a 95 percent bootstrap confidence interval of 0.9014 to 0.9896, followed closely by all-cause mortality at 0.9507 with a 95 percent confidence interval of 0.9040 to 0.9882. The model demonstrated strong discrimination for wound cure, achieving an area under the curve of 0.8800 with a 95 percent confidence interval of 0.8082 to 0.9425, whereas discrimination for recurrence reached 0.8202 with a 95 percent confidence interval of 0.7220 to 0.9109, improving upon the lower recurrence area under the curve of 0.6064 observed during internal development cross-validation.

To provide model interpretability for clinical practice, post-hoc Shapley additive explanations (SHAP) were calculated exclusively for the neural network branch in pre-sigmoid logit space. Global importance rankings and beeswarm distributions revealed distinct shared and outcome-specific predictor patterns across the clinical endpoints. White blood cell count and lymphocyte percentage emerged as consistent inflammatory markers across all four outcomes, exhibiting pronounced nonlinear attribution thresholds at approximately 10 times 10 to the ninth power cells per litre for white blood cells and 20 percent for lymphocyte fraction. Higher white blood cell counts shifted log-odds toward unfavourable prognostic events, whereas elevated lymphocyte percentages correlated with favourable trajectories. In addition, peripheral arterial disease and peripheral neuropathy contributed prominently to amputation predictions, claudication history strongly influenced recurrence risk, and hypertension ranked among the foremost contributors to both amputation and mortality. Metabolic markers also displayed complex associations, including non-linear lipid patterns and elevated fasting and postprandial glucose levels correlating with impaired healing and elevated mortality risks.

Translating the locked ensemble into clinical practice, the authors deployed the software within a web-based outpatient interface and executed an exploratory prospective randomised comparative evaluation across three primary care hospitals from January to March 2026. A total of 202 adults presenting with confirmed diabetic foot ulcers were randomly allocated in a 1 to 1 ratio to receive either routine standard-of-care management or artificial intelligence-assisted care, in which treating clinicians drafted an initial plan, inspected algorithmic probability forecasts alongside SHAP attribution charts and automated decision prompts, and adjusted clinical orders accordingly. Following the active withdrawal of two participants, 200 patients entered the modified intention-to-treat population (100 per study arm). Operational efficiency logging confirmed that clinical implementation imposed modest logistical demands: median patient-level data entry required 4.2 minutes (interquartile range 3.6 to 5.1 minutes), algorithmic probability generation took 2.3 seconds (interquartile range 1.9 to 2.8 seconds), and clinician recommendation review required 2.1 minutes (interquartile range 1.6 to 2.8 minutes).

Assessment of recommendation quality was conducted through a nested paired study of the 100 usual-care cases, for which AI-assisted versions were generated offline and compared blindly by a multidisciplinary panel of four endocrinology, burn, and plastic surgery specialists across three structured dimensions: inappropriate content, omitted content, and scientific soundness. Overall recommendation quality scores increased significantly from 6.47 plus or minus 1.21 points under standard clinician charting to 7.77 plus or minus 1.06 points with algorithmic support, demonstrating an unadjusted mean improvement of 1.30 points with a 95 percent confidence interval of 1.08 to 1.52 (p < 0.0001). A linear mixed-effects model adjusting for reviewer variations yielded an identical mean difference of 1.30 points, with the greatest individual domain gain observed in recommendation completeness (adjusted mean difference 0.57, 95 percent confidence interval 0.43 to 0.71), indicating that automated guidance primarily assisted clinicians in systematically documenting essential follow-up schedules, warning signs, glycemic management plans, and referral triggers.

Prospective patient-level follow-up revealed that artificial intelligence-assisted care was associated with superior short-term care engagement relative to usual care. Patients in the assisted cohort demonstrated significantly higher 4-week follow-up response and revisit rates, accompanied by superior adherence to prescribed glucose monitoring regimens and medical treatments, with directional advantages persisting at 12 weeks across primary and sensitivity analyses. However, the study authors explicitly emphasised that these prospective findings represent improvements in outpatient recommendation completeness and short-term behavioural engagement rather than proven therapeutic efficacy. Over the brief twelve-week prospective observation window, the evaluation was neither designed nor powered to establish definitive reductions in ulcer recurrence, lower-extremity amputation, or all-cause mortality, leaving downstream impacts on hard clinical endpoints unconfirmed.

The authors highlighted several critical methodological limitations inherent to the study. Retrospective derivation from electronic health records introduces potential information and selection biases, and several established clinical prognostic indicators—such as anatomical ulcer depth staging by Wagner grade, ankle-brachial index measurements, and dedicated lower-extremity vascular imaging—were unavailable across the historical records. Furthermore, geographical derivation was restricted to three regional hospital systems in eastern China, precluding assessment across diverse international healthcare systems or distinct racial and ethnic cohorts. The four prognostic endpoints were formulated as independent binary classification tasks across available follow-up, which does not explicitly account for time-to-event dependencies or competing survival risks, such as mortality precluding subsequent ulcer recurrence. Finally, while the web platform incorporated a QR code feedback mechanism to gather longitudinal data for potential offline updates, no model updating was executed during the trial, necessitating future prospective validation before algorithmic retraining or clinical deployment.

Subject of Research: Explainable artificial intelligence prognostic assessment and decision support in diabetic foot ulcer care

Article Title: An explainable artificial intelligence system for prognostic assessment and decision support in diabetic foot ulcer care: a multicentre derivation and validation study with prospective randomised evaluation

Article References: Zhang, J.-Y., Zhuang, Y., Cao, Q., Song, C., Feng, Y., Zhuang, G., Ji, Z., Li, X., Li, Q., Jin, P., Zhang, A., & Zhang, C. (2026). An explainable artificial intelligence system for prognostic assessment and decision support in diabetic foot ulcer care: a multicentre derivation and validation study with prospective randomised evaluation. eClinicalMedicine, 101, Article 104284. https://doi.org/10.1016/j.eclinm.2026.104284

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104284

Keywords: artificial intelligence, diabetic foot ulcers, machine learning, prognostic modeling, clinical decision support, SHAP analysis, amputation prevention, randomised controlled trial, explainable, artificial, intelligence, system