Artificial intelligence has expanded rapidly across podiatric care, from automated gait analysis to the prediction of diabetic foot complications. This narrative review synthesizes the evidence published between 2016 and 2025 on the applications, reported performance, and limitations of these tools. A documented search into five thematic axes: automated gait analysis and biomechanical assessment; diabetic foot risk prediction and ulcer assessment; diagnostic imaging and foot deformity classification; wearable sensing and remote monitoring; and validation, bias, and clinical translation. Deep learning reconstructs plantar pressure and joint kinetics from wearable signals and automates radiographic measurements of foot deformities. Prediction models for diabetic foot ulcers, amputation, and recurrence report areas under the curve between 0.89 and 0.97 in internal validation, although performance attenuates on external testing. Wearable insoles and sensor-enabled dressings extend monitoring beyond the clinic. Convergent evidence indicates that most models remain single-center proof-of-concept studies without prospective deployment, and that podiatrists favor supportive rather than autonomous roles for artificial intelligence. Future research should prioritize external validation, fairness, and implementation studies that link these tools to clinical outcomes.
La inteligencia artificial se ha expandido con rapidez en la atención podológica, desde el análisis automatizado de la marcha hasta la predicción de complicaciones del pie diabético. Esta revisión narrativa sintetiza sobre las aplicaciones, el rendimiento reportado y las limitaciones de estas herramientas. Una búsqueda documentada en Scopus, OpenAlex, PubMed y Crossref identificó estudios organizados en cinco ejes temáticos: análisis automatizado de la marcha y evaluación biomecánica; predicción del riesgo de pie diabético y valoración de úlceras; imagen diagnóstica y clasificación de deformidades del pie; sensores vestibles y monitorización remota; y validación, sesgo y traslación clínica. El aprendizaje profundo reconstruye la presión plantar y la cinética articular a partir de señales vestibles y automatiza las mediciones radiográficas de las deformidades del pie. Los modelos de predicción de úlceras de pie diabético, amputación y recurrencia reportan áreas bajo la curva entre 0,89 y 0,97 en la validación interna, aunque el rendimiento se atenúa en la prueba externa. Las plantillas instrumentadas y los apósitos con sensores extienden la monitorización más allá de la consulta. La evidencia convergente indica que la mayoría de los modelos siguen siendo pruebas de concepto unicéntricas sin despliegue prospectivo, y que los podólogos prefieren roles de apoyo antes que autónomos para la inteligencia artificial. La investigación futura debe priorizar la validación externa, la equidad y los estudios de implementación que vinculen estas herramientas con desenlaces clínicos.
Diabetic foot disease and disorders of gait and foot structure concentrate a large share of preventable disability in adult populations. A meta-analysis of the global epidemiology of diabetic foot ulceration reported a pooled prevalence of 6.3% (95% CI 5.4 to 7.3), with the highest regional estimate in North America at 13.0%.(1) Foot ulceration is the most common lower-extremity complication of diabetes, and recurrence after healing remains common even with structured prevention.(2) The scale of the problem is considerable: it has been estimated that a person with diabetes develops a foot ulcer every second and that at least half of these wounds become infected.(3) Gait impairment and fall risk extend this burden beyond diabetes to older adults, in whom foot-mounted inertial sensors have resolved distinct walking phenotypes with fall prevalence ranging from 27% to 68%.(4) For the podiatrist, the clinical workload spans repetitive measurement of angles and arch height, interpretation of plantar pressure and gait, and risk stratification of the diabetic foot, all of which are tasks that pattern recognition can support.
Artificial intelligence entered clinical medicine through risk assessment and image analysis, and its reach broadened substantially once deep learning matured in the early 2000s.(5) Machine learning differs from conventional statistical modeling in its emphasis on prediction rather than inference and in its capacity to learn directly from high-dimensional signals such as images and sensor time series.(6) In diabetes care, these methods have been proposed for screening and diagnostics, including cloud-based prediction from routine clinical variables such as glucose, body mass index, and age,(7) and reviews of artificial intelligence in diabetes describe gains in diagnostic accuracy alongside unresolved challenges in deployment.(8) The same techniques apply directly to the foot, where the signals of interest, plantar pressure, gait kinematics, wound photographs, and radiographs, are abundant and increasingly digitized, which makes podiatry well suited to clinical machine learning.
Within podiatry the literature is fragmented by task and by modality. Reviews have synthesized image-based diagnosis of diabetic foot ulcers,(9) thermography of the diabetic foot,(10) smart wearable technology for the diabetic foot,(11) digital health technologies for ulcer care,(12) and computational methods for chronic wounds.(13) Each of these addresses a single technique or condition, and none integrates the span from gait analysis to risk prediction or examines in one place the recurring problems of external validation, bias, and clinical translation. The result is that a clinician seeking an overview of what artificial intelligence can already do, and what it cannot yet be trusted to do, finds no single synthesis to consult. The most recent wave of studies, published in 2024 and 2025, has moved from single-center proof of concept toward multicenter prediction and portable measurement, which makes the present moment the point at which a critical synthesis of the state of the field is most useful to clinicians.
The objective of this review is to synthesize the evidence on the applications of artificial intelligence in podiatry published between 2016 and 2025 and to characterize their methodological strengths, limitations, and barriers to clinical translation. Specifically, the review describes the main approaches used in gait analysis and biomechanical assessment; examines the reported performance and clinical validity of models for diabetic foot risk prediction and ulcer assessment; identifies applications in diagnostic imaging, wound care, and remote monitoring; and characterizes the validation gaps and translation barriers that separate these tools from routine practice.
The synthesis included 65 studies distributed across five thematic axes (Table 1). The axes move along the clinical pathway, from the biomechanical signals of gait and plantar pressure to the prediction of diabetic foot outcomes, and close with the methodological conditions that determine whether any of these tools reaches practice.
Table 1. Synthesis of artificial intelligence applications in podiatry by thematic axis.
Axis | Core tasks | Key finding | Studies (n) |
|---|---|---|---|
Gait analysis and biomechanics | Plantar pressure, ground reaction forces, joint kinetics, arch assessment | Deep learning reconstructs kinetics from wearable signals and automates arch assessment | 11 |
Diabetic foot risk and ulcer assessment | Ulcer classification, amputation, recurrence, neuropathy prediction | High internal discrimination, attenuated on external validation | 19 |
Diagnostic imaging and deformity classification | Radiographic angles, landmarks, arch type | Automated measurements comparable to expert reading | 13 |
Wearable sensing and remote monitoring | Insoles, smart dressings, fall detection | Feasibility demonstrated, clinical effectiveness undefined | 8 |
Validation, bias, and translation | External validation, fairness, explainability, language models | Scarce external validation; trust and bias unresolved | 14 |
Note: The number of studies reflects the primary axis to which each study was assigned in the argument matrix.
Plantar pressure and forceplate recordings are the dominant data sources for automated foot and gait assessment, and deep learning is increasingly used to estimate kinetic quantities that previously required laboratory equipment. A dataset of high-resolution plantar pressures collected across footwear conditions and walking speeds illustrates the shift toward open, reusable resources for model development.(15) From single forceplate signals, a physics-informed deep network has been proposed to predict bilateral ground reaction forces and center of pressure, which reduces the need for patients to target separate plates during walking trials.(16) Similar objectives are pursued from wearable inertial sensors: cross-modal learning can infer plantar pressure distribution, ground reaction force, and center of pressure from inertial measurement units,(17) and a hybrid convolutional recurrent network estimates dynamic plantar pressure distributions from wearable signals as a portable alternative to laboratory platforms.(18) Ankle joint moments, which require inverse dynamics modeling in a laboratory, have also been estimated from pressure insoles through a combination of genetic-algorithm sensor selection and deep forest regression (19) (Table 2).
Table 2. Representative models for automated gait and biomechanical assessment.
Study | Task | Data source | Approach |
|---|---|---|---|
Physics-informed deep learning (16) | Bilateral GRF and CoP | Single forceplate | Physics-informed residual network |
Motion2Press (17) | Plantar pressure, GRF, CoP | IMU | Cross-modal deep learning |
Hybrid CNN-BiLSTM (18) | Dynamic plantar pressure | IMU | Convolutional recurrent network |
Ankle moment estimation (19) | Ankle joint moment | Pressure insoles | Genetic algorithm plus deep forest |
Multicenter segmentation (20) | Plantar pressure analysis | Pressure platforms | Machine learning segmentation |
Arch index estimation (21) | Foot arch index | Plantar pressure | Deep learning system |
Flatfoot detection insole (22) | Flatfoot screening | Triboelectric insoles | Self-powered sensing plus learning |
Note: GRF = ground reaction force; CoP = center of pressure; IMU = inertial measurement unit.
A second line of work automates the analysis of plantar pressure images themselves. Conventional proportion-based segmentation of the foot into regions is limited for atypical foot structures, and machine learning segmentation trained across multicenter data has been advanced as a more robust alternative.(20) The same pressure images support arch assessment: a deep learning plantar pressure system estimates the foot arch index for static evaluation of flatfoot and high arch,(21) and a self-powered triboelectric insole detects flatfoot during gait.(22) These tools pursue the same objective: replacing subjective or time-consuming clinical measurement of the medial longitudinal arch with reproducible digital estimates.
Gait classification extends beyond kinetics to diagnostic support. Force sensors combined with machine learning have been used for severity classification in Parkinson disease,(23) and pressure-sensing insoles that capture center-of-pressure features during the timed up and go test distinguished Parkinson disease from controls using 144 features in a study of 77 participants.(24) A flexible triboelectric pressure array combined with deep learning provides a soft, self-powered gait detection system intended for continuous health assessment.(25) These systems resolve foot-specific kinetic, arch, and balance quantities with reported accuracy comparable to reference instruments, although none of them has been evaluated in routine podiatric practice. The convergence of the axis is technical maturity; the divergence is between laboratory equivalence and clinical deployment, and no study in this axis reported a prospective change in podiatric decision making.
A further convergence is methodological. The field is moving from proprietary, single-laboratory data toward shared resources and multicenter benchmarks, as the open plantar pressure dataset
The largest body of work addresses the diabetic foot, where the clinical stakes are highest. Image-based deep learning models classify diabetic foot ulcers from photographs, and explainability frameworks are increasingly added to support clinical trust: a feature-explainability framework was proposed for ulcer identification,
Clinical prediction models built on routine variables target amputation, the most severe outcome. In 598 patients with diabetic foot ulcers, a gradient boosting machine using 17 selected features predicted major amputation with an area under the curve of 0.95 and accuracy of 0.94, with multidrug-resistant infection, C-reactive protein, and diabetes duration as the leading predictors.(28) A support vector machine trained on 149 patients reached an area under the curve of 0.89, and identified C-reactive protein and Wagner grade as independent predictors of amputation.(29) A larger model of 1,332 patients reported an internal area under the curve of 0.93 with XGBoost, which fell to 0.83 on external validation in a separate multicenter cohort, a pattern that illustrates the gap between internal and external performance.(30) A review of risk prediction models for the diabetic foot confirms this rapid growth and the same recurring limitation (31) (Table 3).
Table 3. Representative models for diabetic foot risk prediction and ulcer assessment.
Study | Outcome | Population | Model and key performance |
|---|---|---|---|
Corneal nerve imaging (32) | Diabetic foot risk | 471 participants | Deep learning plus HbA1c, AUC 0.94 |
Major amputation (28) | Amputation | 598 patients | Gradient boosting, AUC 0.95 |
Multicenter prediction (33) | Incident diabetic foot | 6,180 older adults | Logistic ensemble, AUC 0.94 |
Infection versus osteomyelitis (34) | Differential diagnosis | 3,612 patients | LightGBM, external AUC 0.94 |
Amputation (30) | Lower-extremity amputation | 1,332 patients | XGBoost, internal AUC 0.93, external 0.83 |
Recurrence (35) | Three-year recurrence | 494 patients | XGBoost, AUROC 0.92 |
Amputation (29) | Amputation | 149 patients | Support vector machine, AUC 0.89 |
Ulcer in arterial occlusion (36) | Incident ulcer | Diabetic patients with occlusion | Clinical model, AUC 0.96 |
Note: AUC = area under the curve; AUROC = area under the receiver operating characteristic curve.
Multicenter development represents a step toward generalizable prediction. A study of 6,180 older adults from 11 community hospitals combined deep and classical learning and reported a logistic regression ensemble with areas under the curve of 0.94 in validation and 0.94 in the test set, with ankle brachial index and lower limb discomfort among the informative features.
Beyond ulcer classification, models target the neuropathy and tissue stress that precede ulceration. Plantar pressure analysis with gradient boosting classified diabetic peripheral neuropathy with 100% accuracy in static assessment of 86 patients, with forefoot-to-hindfoot pressure ratio as a key feature.
The axis also exposes a definitional divergence that affects comparability. The target outcome ranges from incident ulceration and ulcer recurrence to healing phase, infection, osteomyelitis, and major amputation, and each study selects its own definition and time window.
Radiographic measurement of foot deformities is repetitive and prone to interobserver variability, which makes it a natural target for automation. A deep learning solution automated weight-bearing foot measurements and was compared against manual reading by radiologists.
Portable measurement extends the same logic beyond the radiology workstation. A smartphone application measured the hallux valgus angle from radiographs displayed on a monitor, with three users of different training levels.
Continuous monitoring moves assessment out of the clinic, and the diabetic foot is the main target. A systematic review of smart wearable technology for diabetic foot prevention, assessment, and management identified sensing and smart miniaturized devices as an expanding field while noting the limited consolidated clinical evidence.
Sensor-enabled dressings close the loop between monitoring and treatment. An in vivo validated smart dressing embeds a pH-sensitive dye in a hydrogel sheet, so that the wound environment signals infection remotely, an approach motivated by the estimate that at least half of diabetic foot ulcers become infected.
Fall risk assessment illustrates the breadth of wearable foot sensing beyond diabetes. Foot-mounted inertial waveforms from 146 older adults were clustered into four gait phenotypes with fall prevalence between 27% and 68%, supporting phenotype-specific prevention.
Across the five axes, the recurring constraint is not algorithmic capacity but evidence of clinical validity. External validation remains the exception: only a minority of the foot models reviewed were tested outside their development center, and where they were, performance dropped, as in the amputation model that fell from an area under the curve of 0.93 internally to 0.83 externally.
Explainability is now reported as a routine companion to predictive accuracy. SHAP, LIME, and Grad-CAM feature in models for infection versus osteomyelitis,
Language models introduce a new channel through which patients and clinicians encounter podiatric information. ChatGPT-3.5 answered questions on diabetic foot ulcer care with high rated accuracy but omitted key details and presented some misinformation in one evaluation,(66) and a retrieval-augmented generative system was built to provide diabetes and foot care information at an eighth-grade literacy level with higher accuracy.(67) The contrast between the two is instructive: a general model answers from its training distribution, whereas a system anchored to retrieved medical knowledge reduces the risk of fabricated content, and the latter approach is the more defensible basis for patient education. Qualitative work with podiatrists indicates a clear preference for artificial intelligence in supportive roles over diagnostic roles, driven by concerns about trust and accountability.(68)
Reviews of the field converge on the same diagnosis. Systematic and narrative syntheses of artificial intelligence in chronic wounds,
The principal finding of this review is that artificial intelligence in podiatry has matured technically across the entire care continuum, from gait and plantar pressure analysis to diabetic foot risk prediction, while the evidence of clinical validity and prospective benefit has not kept pace. Across 65 synthesized studies, models report high internal discrimination, several with areas under the curve above 0.9, but external validation remains the exception and deployment studies are nearly absent. This pattern is the same one described in the wider medical literature, where machine learning outperforms conventional statistics in prediction tasks while wide clinical adoption lags.
The findings align with and extend prior reviews. Syntheses of image-based diabetic foot diagnosis,
Several explanations account for the divergence between internal accuracy and clinical readiness. Most studies are retrospective and single-center, and their samples, ranging from 86 to 6,180 participants, are convenience samples drawn from one or a few hospitals.(33,38) Predictor sets differ across models, which complicates comparison and replication, and the outcome definitions for amputation and recurrence vary between studies.(28,35) Sensor-based systems introduce additional heterogeneity through hardware, sampling, and preprocessing choices.
The mechanism behind external attenuation deserves emphasis, because it is not a defect of the algorithms but a property of the data. A model optimized on one hospital's case-mix learns that institution's prevalence, coding practices, and referral thresholds, and when these shift, the same coefficients no longer separate classes as cleanly. This is the dataset shift described for artificial intelligence in medicine at large,
For podiatric practice the implications are specific. Automation of radiographic measurements and plantar pressure analysis is close to usable in the clinic, because these tasks replace repetitive measurement rather than clinical judgment, and the reliability evidence is consistent.(20,43) Risk prediction is less ready for autonomous use, and the preference of podiatrists for supportive roles is consistent with the appropriate deployment of these models as triage and documentation aids rather than decision makers.
For the clinician, a practical reading of this literature is a three-part rule. Measurement automation, for angles and arch parameters, is reliable enough to adopt where validated software is available.
These tools also carry consequences for podiatric education and governance. The qualitative finding that podiatrists prefer supportive roles for artificial intelligence
These observations translate into concrete design standards. Future studies should report both discrimination and calibration, specify the outcome definition and its time window, and describe the population from which the sample was drawn well enough to judge transportability.(29,35) Models intended for screening should be tested prospectively against the clinical pathway they are meant to replace, and models intended for monitoring should be compared against periodic examination, not against no care at all. Benchmark datasets and shared evaluation protocols would let independent groups reproduce results, which the current single-center literature does not yet permit.(15,20) These standards cost less than the effort already invested in model development, and they are the shortest path from demonstration to deployment.
The review has limitations. First, the selection was narrative rather than systematic, so the study set reflects the documented searches and the thematic matrix rather than an exhaustive screen; the risk of missing relevant work is real, and the remedy is to repeat the search in living systematic reviews as the field grows.
The research agenda that follows is concrete. External validation should become a condition for reporting foot and ankle models, and multicenter, prospective cohorts with harmonized outcome definitions should replace single-center retrospective designs.
Table 4. Limitations of the current evidence and corresponding recommendations.
Limitation | Recommendation |
|---|---|
Single-center retrospective designs | Prospective multicenter cohorts with harmonized outcome definitions |
Scarce external validation | Require external validation as a reporting condition |
Bias in ulcer classification | Report fairness across skin tone, sex, and age |
No evidence of clinical effect | Implementation studies measuring triage, healing, and amputation rates |
Heterogeneous sensors and preprocessing | Standardized data and benchmark datasets |
Access inequities in remote monitoring | Usability and equity studies in low-resource settings |
Note: The recommendations are linked to each limitation as proposed in the discussion. Source: Own elaboration.
Automated gait analysis and biomechanical assessment can reconstruct plantar pressure and joint kinetics from wearable signals and can measure the foot arch with accuracy comparable to manual methods, although the evidence is mostly single-center proof of concept. Models for diabetic foot risk prediction and ulcer assessment report high internal discrimination, with areas under the curve between 0.89 and 0.97, and their performance attenuates on external validation, which currently limits clinical deployment. Diagnostic imaging and deformity classification, wearable sensing, and remote monitoring extend artificial intelligence across the podiatric care continuum but lack prospective evidence of benefit. The central barriers are scarce external validation, unresolved bias, and the absence of implementation studies, which define the agenda for the next stage of the field.
No financing
None
Drafting – original draft: Jose Ignacio Robaina Castillo.
Writing–review and editing: Jose Ignacio Robaina Castillo.