Combining thermography and artificial intelligence in comparison with a diabetic foot nurse for diabetic foot ulcer detection: A diagnostic accuracy study

Abstract

Background

Early detection of diabetic foot complications is essential to prevent ulcers and amputations. Thermographic imaging offers a non-invasive method for identifying risk, but clinical interpretation has traditionally relied on human thermographers. Artificial intelligence (AI) may offer a more scalable and objective alternative.

Objective

To evaluate the diagnostic performance of an AI-powered thermographic screening tool in identifying risk for diabetic foot complications, compared to nurse-led clinical assessment.

Methods

We conducted a cross-sectional study of 100 adults with diabetes undergoing routine foot screening. For each participant, a smartphone-based thermal imaging device was first used to capture plantar images, from which the AI model generated risk scores (0–3). Second, a diabetic foot nurse performed a clinical examination and assigned the reference risk scores (0–3). Absolute temperature differences were computed from thermal images, and diagnostic accuracy metrics were calculated using the nurse assessment as the reference standard.

Results

The AI system demonstrated 100% sensitivity, 96.8% specificity, 66.7% positive predictive value, and 100% negative predictive value for detecting moderate-to-high risk cases. There was a strong correlation between AI and nurse scores (ρ = 0.973), and both assessors showed increasing temperature asymmetry with higher risk levels.

Conclusions

The AI model accurately detected all moderate-to-high risk cases flagged by the nurse, with high sensitivity and specificity. Its strong alignment with thermal data and consistent scoring suggest its value as a scalable and reproducible adjunct for diabetic foot screening. Further validation in longitudinal settings may support broader integration in remote and primary care environments.

Introduction

Diabetes mellitus

Diabetes affects 1 in 10 adults worldwide (537 million). ¹ This number is predicted to rise to 643 million by 2030 and 783 million by 2045. In the Middle East and North Africa (MENA) region, the prevalence is higher as it affects 1 in 6 adults (73 million). ¹ Despite advances in medical therapies, the prevalence of diabetes mellitus and diabetes-related complications continues to rise.

One of the most common complications of diabetes is diabetic foot ulcer (DFU). It is estimated that one-third of people with diabetes will develop a DFU during their lifetime. ² Unfortunately, even after a DFU has been resolved, recurrence is common and is estimated to be 40% within one year, 60% within three years, and 65% within five years. ² Lower limb amputation is the most severe and costly outcome if DFU complications persist. ³ Furthermore, diabetes foot care costs are the single largest category of diabetes-related medical costs. A study by Armstrong et al. estimated that diabetic foot care accounts for one-third of all diabetes-related costs. ⁴ Prevention of these lower limb complications could lead to a significant positive impact on health outcomes and significant cost savings. Unfortunately, current tools for detecting DFU have limited scalability in terms of time efficiency and practicality. ⁵

Traditional methods for detecting diabetic foot in the modern era

In the diagnosis of DFUs, clinicians have long depended on traditional methods, each bearing distinct advantages and limitations. The following section critically examines these diagnostic tools.

Monofilament test

The monofilament test is not always reliable for early detection of DFUs because it primarily identifies loss of protective sensation (LOPS), missing other risk factors like arterial insufficiency and structural deformities.^6,7 The test's accuracy can also be influenced by the technique of the administrator, patient factors like swelling and temperature, and may not pick up early or intermittent neuropathy symptoms. ⁷ For a more comprehensive assessment, clinicians often use the monofilament test in conjunction with other evaluations and tests.

Ankle-brachial index

Foot ulcers can be exacerbated by underlying peripheral arterial disease (PAD). The ankle-brachial index (ABI) is a widely used, non-invasive method for diagnosing PAD. However, its accuracy is debated, especially in patients with exertional leg pain. ⁸ Furthermore, it is not useful to detect PAD among patients with diabetes because calcified vessels can distort results.^9,10 This raises concerns about the potential for an unreliable ABI to overlook those at heightened risk for foot ulcers, emphasizing the importance of a comprehensive assessment in diagnosing PAD and determining ulcer risk.

Duplex ultrasonography

It is a common imaging test used to diagnose and monitor PAD. It can accurately identify the location and severity of narrowing or blockage in arteries. However, some artery segments are difficult to visualize with duplex ultrasonography (DUS), especially in the lower limb, and the results may not be fully reliable. ¹¹ Additionally, DUS is not well-suited for people with diabetes because their calcified vessels are less flexible and more difficult to see with ultrasound. ¹² For this reason, other testing methods are often more reliable for diagnosing PAD in people with diabetes.

Angiography

The most accurate way to diagnose DFUs is by using a special type of X-ray called an angiogram; however, there are some drawbacks to this test. ¹³ First, repeated exposure to X-rays can be harmful to patients’ health. Second, the test is expensive and time-consuming, and sometimes patients need to be hospitalized overnight for the procedure.

Routine clinical foot assessment

Routine foot screening remains the cornerstone of diabetic foot assessment in modern clinical practice and serves as the standard of care. ¹⁴ International guidelines, such as the International Working Group on the Diabetic Foot (IWGDF), recommend that every person with diabetes undergo an annual foot examination. ¹⁴ During these examinations, clinicians inspect the foot and assess pedal pulses and protective sensation. Risk status is then classified into categories such as low, increased, or high risk to guide follow-up intervals. ¹⁴ However, despite its widespread use, this examination is manual, episodic, and highly dependent on the examiner's expertise. It varies significantly between providers and clinical settings, and subclinical abnormalities, such as localized inflammation, tissue stress, or perfusion abnormalities, may go undetected between visits. ¹⁵

Human medical thermography

Human medical thermography results from decades of research and development in the performance of infrared imaging equipment, standardization of technique, and clinical protocols for thermal imaging.^16,17 It could visualize diseases not readily detected or monitored by other methods. It is a fast, passive, non-contact, and non-invasive imaging method that has been used by numerous peer-reviewed studies. ¹⁸ It is currently used globally to screen, detect, and monitor diseases. The American Academy of Thermology (AAT) established guidelines for the use of thermography in the evaluation of diabetic patients. ¹⁹ These guidelines provide recommendations for the use of thermal imaging in the detection and monitoring of diabetic neuropathy, including protocols for image acquisition and interpretation. ¹⁹ Figure 1 shows the thermal images of the lower extremities of two diabetic subjects.

Figure 1.

Typical thermal images of plantar region for healthy, at risk, and amputated subjects.

Human Medical Thermography has many advantages that could encourage widespread adoption. Thermal imaging is relatively inexpensive, compact, portable, involves no ionizing radiation, and requires little electric power. Recent technological breakthroughs have transformed large and expensive stationary cameras into portable and inexpensive solutions while maintaining quality imaging. ²⁰

A major limitation of the current state of medical human thermography is that even the most skilled human thermographer can only observe, analyze, and successfully interpret a limited number of thermograms. Computers, however, can process an image efficiently and extract useful information without tiring. Leveraging artificial intelligence (AI) algorithms, specifically, computer vision, can objectively observe the findings and minimize inter-observer variability. Ongoing progress in software image analysis and reduced reliance on human labor results in faster throughput and centralized processing. This can lead to increased thermographic accuracy and reliability. Nevertheless, computer-aided thermography will require high-level training and experience to ensure quality outcomes.

Diabetic foot ulcers and thermography

Thermography is helpful for the early detection of abnormalities of the foot by analyzing asymmetries and local temperature changes over time. Assessing temperature differences can enable the early detection of ulcers.^21,22 Peripheral vascular disease (PVD) is a common complication of diabetes, which can result in alterations in blood flow that induce changes in skin temperature. These changes in skin temperature may also indicate tissue damage or inflammation resulting from trauma or excessive pressure. The etiology of these traumas is frequently related to moderate repetitive stress that goes unnoticed due to diabetic neuropathy. The application of thermal imaging for the detection of diabetic foot complications is based on the premise that variations in plantar temperature are associated with these types of complications.^5,22–29 Furthermore, there appears to be a positive correlation between body mass index (BMI) and the risk of diabetic foot complications in patients with type 2 diabetes.^30,31

The rapid development of handheld smartphone-based thermal infrared imagers presents a creative solution for detecting and monitoring DFUs. ³² To address the lack of thermographers, practical computer vision algorithms are needed to automate the process of image acquisition and analysis. These rapidly expanding, low-cost, and widely available resources can help predict one's risk of developing foot ulcers, potentially saving limbs and lives.

AI and its applications are increasingly demonstrating promise in the detection and management of DFUs.^33,34 Diabetes foot syndrome, with its lack of early symptoms and significant impact on patients’ quality of life, necessitates the use of AI in timely screening and detection of risk for foot ulcers and possible amputations.^33,34

Studies such as Peregrina-Barreto et al.'s showed the potential of infrared thermography for detecting foot complications in diabetic patients. ³⁵ Several researchers demonstrated promising results in the detection of DFUs using machine learning techniques to analyze foot images.^36–38 These findings suggest that AI's application to data derived from thermal plantar images yields promising results. AI has a significant potential to revolutionize the detection and management of DFUs. In this study, we will leverage AI technologies that are deployed on a smartphone-based thermal imager and application.

The identification of DFUs using thermogram images combined with AI is not an area that has been extensively researched. Our proposed system has been developed using advancements in thermography and AI techniques.

Methods

Study design and objective

This cross-sectional diagnostic accuracy study followed the Standards for Reporting of Diagnostic Accuracy Studies (STARD) guidelines. ³⁹ The study was conducted to evaluate the performance of TFScan, a smartphone-based thermal imaging system integrated with AI, in assessing the risk of DFUs. TFScan's risk results were compared to the International Working Group on the Diabetic Foot (IWGDF) risk stratification categories assigned by an experienced diabetic foot nurse, which served as the reference standard. ¹⁴ This reference standard was selected because clinical examination by an experienced diabetic foot healthcare worker is the standard of care for routine risk stratification in real-world practice.^13,14 More definitive alternatives, such as angiography, were not feasible or ethical for our study.

Thermal imaging system

The study employed TFScan, a commercially available, handheld infrared thermal imager integrated with a smartphone interface and custom-developed software for medical-grade thermal imaging. It enabled real-time acquisition, processing, and classification of plantar foot images. Figure 2 shows the thermal imaging system TFScan.

Figure 2.

Thermal imaging system, TFScan.

The imaging system comprised three main components: (1) a portable thermal sensor, developed by Teledyne FLIR LLC; (2) a smartphone for data capture and interaction; and (3) image analysis software leveraging computer vision algorithms, developed by Amplifai Health LTD. The analysis software applies computer-vision-based algorithms to extract quantitative thermal features, including absolute temperature asymmetry, regional temperature variations, and abnormal thermal distribution patterns. The system generates a risk output mapped to four ordinal categories (0–3) based on predefined thresholds, including an asymmetry cutoff of  ≥ 2.2 °C, following the AAT guidelines. ¹⁹ This automated approach aimed to minimize inter-observer variability and increase diagnostic consistency.

Study setting and recruitment

Participants were recruited from Johns Hopkins Aramco Healthcare (JHAH) and formed a consecutive series. Eligibility was determined through a review of electronic medical records, including recent HbA1c levels, fasting blood glucose, laboratory investigations, current diabetes medications, and annual screening outcomes.

Eligible participants were adults aged 18 years and older, including those diagnosed with diabetes and healthy individuals with no known history of diabetes or cardiovascular disease. Exclusion criteria included active diabetic foot complications (e.g. visible ulcers, infections, wounds, cellulitis, or amputations), the need for debridement, or the inability to stand without assistance due to increased fall risk.

All participants provided written informed consent prior to enrollment. Following consent, the case manager coordinated the assessment session, and participants were positioned comfortably in bed for the imaging procedure.

Ethical considerations

The study received ethical approval from the Institutional Review Board at JHAH (IRB# 24-24). Participants were informed of the study objectives, procedures, risks, and their rights. Written informed consent was obtained from all participants, and data confidentiality was maintained using unique coded identifiers and secure, password-protected systems.

Data collection

Data were collected between January 2024 and May 2024. Participants completed a baseline questionnaire capturing demographic and clinical data, including age, sex, height, weight, BMI, diabetes duration, HbA1c levels, cardiovascular comorbidities, retinopathy, neuropathy, and PAD. Symptoms such as unsteadiness, burning, tingling, and numbness were also recorded.

A trained nurse conducted the first part of the data collection, including taking the thermal images. To ensure consistency, thermal imaging followed protocols established by the AAT. ¹⁹ Environmental conditions were controlled, maintaining a room temperature between 19 °C and 25 °C with relative humidity below 70%. Participants rested with legs hanging freely for 10 min prior to imaging. The thermal camera was positioned approximately 1 m from the plantar surface of the foot.

A second, experienced diabetic foot nurse independently performed the clinical examination. This assessment followed IWGDF screening recommendations and included a visual inspection of the patient feet, evaluation for LOPS and PAD through palpation of pedal pulses and identification of clinical signs of vascular insufficiency. The clinical examination informed the nurse-assigned IWGDF risk score.

Importantly, the nurse conducting the thermal imaging, and the diabetic foot nurse assigning the clinical risk score, were blinded to each other, and the diabetic foot nurse was fully blinded to the AI-generated results.

Screening and risk classification

Risk assessments were conducted using the International Working Group on the Diabetic Foot (IWGDF) guidelines, assigning scores from 0 (very low risk) to 3 (high risk). ¹⁴ Both the AI system and the diabetic foot nurse independently assigned scores based on thermal and clinical findings, respectively.

Two analytical approaches were employed. First, full ordinal scores (0–3) were compared to assess agreement across risk levels. Second, scores were collapsed into binary risk categories: low risk (0–1) and high risk (2–3), to facilitate diagnostic accuracy analyses using the nurse's assessment as the reference.

Outcome measures

The primary outcome was the level of agreement between AI-generated and nurse-assigned risk scores. Agreement was evaluated using Cohen's weighted kappa. Secondary outcomes included diagnostic accuracy metrics—sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV)—based on the binary risk classification. Additionally, absolute temperature asymmetry (≥2.2 °C) between the left and right feet was assessed as a surrogate marker for abnormal thermal findings.

Statistical analysis

All analyses were performed using R software (version 4.3.0). Continuous variables were summarized using means and standard deviations, while categorical variables were presented as frequencies and percentages.

The sample size was informed by previously published thermography or diabetic foot diagnostic accuracy studies, therefore, a target of 100 participants was considered appropriate for this validation and captures a meaningful number of high-risk cases based on expected prevalence in similar routine screening.^40–42

Ordinal agreement between AI and nurse scores was assessed using weighted kappa. The Kruskal-Wallis test evaluated differences in temperature asymmetry across ordinal risk levels. For binary comparisons, diagnostic accuracy measures (sensitivity, specificity, PPV, NPV) were calculated with 95% confidence intervals. Chi-square tests assessed associations between categorical classifications. Independent t-tests and Cohen's d effect sizes compared mean temperature asymmetry between low- and high-risk groups. Spearman's rank correlation was used to explore associations between thermal asymmetry and both AI and nurse-assigned scores.

Results

Patient demographics

A total of 100 participants were included in the study. The mean age of participants was 57.26 years (SD = 10.26), with a mean height of 168.41 cm (SD = 7.47) and a mean weight of 83.33 kg (SD = 14.20), yielding a mean BMI of 29.32 kg/m² (SD = 4.31). The majority of the participants were male (77%), with 23% being female.

Clinical history

The average duration since diabetes diagnosis was 11.49 years (SD = 8.39), and over half of the participants (55%) had diabetes for more than 10 years. Poor glycemic control was identified in 32% of participants. Cardiovascular disease was reported in 10% of the sample, while PAD and retinopathy were present in 2% and 3%, respectively. Only 1% of participants reported having been diagnosed with neuropathy.

Physical screening

Among the cohort, 35% presented with at least one abnormal physical finding. Burning sensations were reported by 18% of participants, prickling by 10%, and numbness by 12%. Unsteadiness in walking was the least common symptom, reported by only 1%.

Screening outcomes

Using the full risk classification system (0 = no risk, 1 = low, 2 = moderate, 3 = high), the AI system (TFScan) classified 9% of patients as being at moderate or high risk, whereas the nurse identified 6% in the same categories. When the scores were collapsed into binary categories for simplicity (low risk: 0–1; high risk: 2–3), the AI flagged 9% of participants as abnormal, compared to 6% flagged by the nurse. This slight discrepancy suggests that the AI was more conservative in its classification, potentially favoring over-inclusion to minimize missed cases. Figure 3 shows the distribution of ordinal risk scores assigned by the AI system and the nurse across four levels (0 = no risk, 1 = low, 2 = moderate, 3 = high). While Table 1 summarizes the binary classifications in a 2 × 2 format for diagnostic accuracy calculations.

Figure 3.

Distribution of ordinal risk classifications by AI and nurse assessments.

Table 1.

2 × 2 contingency table comparing TFScan binary risk classification with the nurse's IWGDF-based reference assessment.

	Nurse high risk	Nurse low risk
TFScan high risk	6	3
TFScan low risk	0	91

Moreover, 61% of participants met at least one clinical risk criterion (e.g. cardiovascular disease, PAD, neuropathy, retinopathy, or diabetes duration >10 years), highlighting that a large portion of the cohort had underlying conditions known to increase the risk of diabetic foot complications. While many participants had clinical risk factors that, according to IWGDF guidelines, warrant at least a low risk classification, both the AI and nurse appropriately identified them as such. This consistency suggests that the AI system is aligned with guideline-based risk stratification and is not underestimating risk in high risk populations, which is an encouraging indicator for its potential use in clinical triage and early intervention.

Abnormal asymmetry scans were identified in 8% of participants. This threshold is clinically meaningful, as temperature asymmetry is an early indicator of inflammation or perfusion issues, which can precede the development of DFUs.

Comparison of AI and nurse risk scores

A Spearman correlation analysis revealed a very strong positive correlation between AI and nurse ordinal scores (ρ = 0.973, p < 0.0001), indicating a high degree of consistency in how both the AI and nurse ranked participants across the risk spectrum. Kruskal-Wallis tests showed statistically significant differences in absolute foot temperature asymmetry across the AI-derived risk categories (H = 15.87, p = 0.0012) and across nurse-derived risk categories (H = 8.32, p = 0.0398). In both cases, temperature asymmetry increased with higher risk scores, providing internal validity for the ordinal classification system: both the nurse and AI assigned higher scores to individuals with more pronounced thermal abnormalities.

When exploring the direct relationship between asymmetry and risk scores, Spearman's correlations showed that temperature asymmetry was modestly correlated with both AI (ρ = 0.269, p = 0.0068) and nurse (ρ = 0.213, p = 0.0335) scores. The slightly stronger correlation for the AI suggests it relied more directly on objective thermal features, whereas the nurse may have considered broader clinical context. Figure 4 boxplot shows the distribution of absolute temperature differences (°C) across ordinal risk categories (0–3) as assessed by the AI model and the nurse. Higher risk scores in both groups are associated with larger temperature asymmetries.

Figure 4.

Absolute temperature differences by risk score and assessor.

Diagnostic accuracy of the AI system

When the AI and nurse scores were dichotomized into low versus high risk categories, and the nurse's evaluation was treated as the reference standard, the diagnostic performance of the AI tool was as follows:

• Sensitivity: 100% (CI: 60.9–100%)

• The AI correctly identified all participants who were classified as moderate or high risk by the nurse, indicating no false negatives.

• Specificity: 96.8% (CI: 91.0–98.9%)

• The AI correctly identified nearly all low-risk participants, with only a small number classified as moderate/high risk when the nurse deemed them low risk.

• PPV: 66.7% (CI: 35.4–87.9%)

• Two-thirds of the participants flagged by the AI as moderate/high risk were also classified as such by the nurse, indicating that the AI produced some false positives.

• NPV: 100% (CI: 96.0–100%)

• All participants classified as low risk by the AI were also classified as low risk by the nurse, confirming the AI's strong reliability in ruling out moderate/high risk.

These results suggest that the AI system is highly sensitive and highly specific, with excellent ability to rule out at-risk individuals and a cautious approach that slightly over-identifies moderate/high-risk cases. This profile is particularly valuable in clinical screening settings where missing true positives is unacceptable, and a conservative triage tool is preferred. The slightly lower PPV reflects a safe bias toward over-identification, which can be acceptable in early-stage screening when followed by clinical confirmation.

Discussion

This study evaluated the performance of AmplifAI Health's thermographic AI screening system (TFScan) in comparison to nurse assessments for detecting early indicators of diabetic foot complications. Using an updated four-level ordinal risk classification and paired clinical/thermographic data, the analysis provides a more nuanced assessment of AI-nurse agreement and alignment with physiological markers.

Overall, the AI system demonstrated perfect sensitivity, correctly identifying all patients classified as moderate-to-high risk by the nurse. It also showed high specificity (96.8%), accurately ruling out nearly all low-risk cases. However, the PPV was 66.7%, indicating that about one-third of those flagged by the AI were not classified as high risk by the nurse. This suggests that the AI may sometimes slightly overestimate risk, favoring caution to avoid missing potential complications. It is a reasonable tradeoff in early screening settings, particularly for populations with a high prevalence of underlying risk factors.

The system showed promising alignment with objective thermographic data. Correlation and Kruskal-Wallis analyses demonstrated that both the AI and nurse risk scores increased with foot temperature asymmetry, reinforcing the clinical validity of the risk classification system. Notably, the AI's risk scores had a stronger correlation with absolute temperature differences (ρ = 0.269) than the nurse scores (ρ = 0.213), suggesting that the AI may be more tightly coupled with quantifiable thermal patterns, whereas the nurse assessments might integrate additional subjective or contextual information.

The very strong correlation between AI and nurse ordinal scores (ρ = 0.973) confirms that the AI closely mirrors the nurse's risk stratification patterns across the full spectrum of scores. While the AI flagged slightly more participants as moderate to high risk, its perfect sensitivity and strong correlation with nurse assessments support its potential use as a reliable triage tool, particularly for identifying individuals who require closer monitoring. The consistency in low risk classifications by both the AI and the nurse also suggests that the system can reinforce clinical confidence in safely ruling out urgent concerns.

From a systems-level perspective, thermographic AI tools could fill a critical gap in diabetic foot screening by providing scalable, non-invasive, and objective assessments that complement in-person evaluations. Compared to conventional modalities like angiography or duplex ultrasonography, which are resource-intensive and impractical for routine monitoring, thermographic imaging can be deployed in primary care and community settings, including remote care environments. Furthermore, integrating this technology into existing telehealth platforms could help mitigate diagnostic limitations in virtual consultations—an increasingly important consideration in chronic disease management.

Early intervention significantly reduces the incidence of foot ulcers and amputations in people with diabetes. ⁴³ Regular diabetic foot exams, such as those recommended annually by the IWGDF, are essential for identifying high-risk conditions. More frequent assessments may be warranted based on individual findings. ¹⁴ High-risk patients should receive specialized diabetic foot care to prevent ulcer progression. ⁴⁴ In this context, thermographic tools could serve as valuable adjuncts to clinical evaluations, enabling earlier and more precise detection of foot complications. Moreover, both pharmacological and non-pharmacological interventions are available to improve blood circulation and reduce complications in diabetic feet, making early diagnosis a cornerstone of effective management.

Another critical area of advancement is telemedicine. Both healthcare providers and patients express concerns about the quality of virtual consultations, particularly for conditions requiring detailed physical examination. Current telemedicine solutions rely on basic microphones and cameras, limiting their ability to support accurate remote diagnosis of DFUs. Incorporating at-home thermography into telemedicine platforms could address this gap by enabling accurate and consistent remote diagnosis and monitoring. Regular monitoring of foot temperature has been shown to reduce the likelihood of disabling complications, such as foot ulcers and lower-limb amputations. ⁴⁵ Thus, integrating thermographic tools into telemedicine could significantly enhance the quality of care for individuals with diabetes.

Comparison with previous work

Previous studies have demonstrated the potential of thermography as a noninvasive method to detect early signs of diabetic foot complications, particularly by identifying temperature asymmetries associated with inflammation, neuropathy, and PVD. For example, two studies reported that localized thermal differences can be used to detect pre-ulcerative conditions, often before clinical signs become apparent.^22,28 These studies, however, relied primarily on human interpretation of thermal images.

More recent work has explored the use of machine learning to automate thermographic analysis. Two studies demonstrated that AI models trained on thermal images could distinguish between healthy and at-risk feet with promising accuracy.^31,32 Similarly, Khandakar et al.'s applied deep learning techniques to thermal foot images and reported effective classification performance, suggesting the feasibility of AI-assisted diagnosis. ³⁸

Despite these advances, many of the existing studies were either retrospective or limited to image classification without real-world clinical comparison. In contrast, our study uniquely evaluates an AI-enabled thermographic tool in a cross-sectional, real-world setting and directly compares its performance to nurse assessments using a four-level ordinal risk classification aligned with international guidelines. ¹⁴

While earlier studies reported good overall accuracy, few provided detailed diagnostic performance metrics such as sensitivity, specificity, or predictive values in direct comparison to clinical evaluations. Our findings contribute new evidence by showing that the AI system achieved high sensitivity and specificity, along with strong correlation to objective thermal asymmetry. This suggests that automated thermographic screening may match or even exceed human consistency in interpreting subtle physiological differences, particularly in the early identification of diabetic foot risk.

Moreover, unlike prior work that often focused on isolated image features, our system integrated temperature asymmetry, pattern recognition, and risk scoring into a structured framework. This supports greater scalability and applicability in real-world screening, particularly in remote or resource-limited settings where access to trained clinicians may be limited.

Strengths and limitations

This study has several strengths that support the validity and applicability of its findings. By employing a cross-sectional design in a real-world clinical screening context, it included a diverse sample of adults with diabetes across a range of risk profiles. The use of a standardized four-level risk classification for both AI and nurse assessments enabled detailed ordinal comparisons, while the inclusion of binary groupings supported diagnostic accuracy analyses. The integration of a smartphone-based thermal imaging device with an AI model for risk stratification reflects a novel and scalable approach to diabetic foot screening, particularly relevant in primary care and remote settings. Imaging protocols aligned with the AAT guidelines further enhanced consistency and quality. The study also benefited from parallel evaluations using both objective thermographic measures, such as absolute temperature asymmetry, and independent nurse assessments, allowing for triangulation of algorithmic and clinical interpretations.

Several limitations must also be acknowledged. While the AI system achieved perfect sensitivity and high specificity, its lower PPV indicates that some flagged cases may not align with nurse assessments, reflecting a conservative bias toward identifying possible abnormalities. This could lead to unnecessary follow-up in some cases, although it also supports safety in early screening. The relatively small number of participants classified as moderate or high risk limited subgroup analyses and may have affected the precision of performance estimates. The use of a single diabetic foot nurse as the reference standard may introduce bias and limits the external validity of the accuracy estimates, as the AI may be reproducing the nurse's judgements rather than detecting true pathology. Additionally, interrater variability was not assessed, which further limits generalizability. In addition, the cross-sectional design restricts conclusions about long-term outcomes, such as ulcer development or response to intervention. Moreover, the study did not include a formal sample power calculation, which may limit the precision of the diagnostic accuracy estimates. Finally, while the AI scores showed stronger correlation with quantitative thermographic features than nurse assessments, the model may not fully capture nuanced clinical factors such as symptom progression or biomechanical risk, which are often considered in human evaluations.

Directions for future research

Future research should build on these findings by expanding sample sizes and including more high risk cases. Longitudinal studies in diverse clinical settings are needed to determine whether AI-detected thermographic abnormalities predict actual ulcer development over time, which would establish prognostic validity and assess the generalizability. Further refinement of the AI model, particularly in incorporating multimodal data such as symptom profiles, comorbidities, or sensorimotor deficits, may enhance sensitivity without sacrificing specificity. Comparative effectiveness studies are also warranted to assess how AI-enhanced thermography performs relative to gold-standard imaging techniques (e.g. Doppler ultrasound or angiography) and traditional risk scores. Additionally, cost-effectiveness evaluations are needed to determine whether AI-based foot screening can reduce healthcare utilization and lower the incidence of hospital admissions or amputations. Research exploring patient trust, usability, and equity of access, particularly in remote or underserved communities, will be essential to support wider clinical adoption.

Conclusion

This study evaluated the performance of an AI powered thermographic tool for diabetic foot screening, demonstrating perfect sensitivity, high specificity, and strong agreement with nurse assessments. The AI model showed a stronger correlation with objective thermal data than the nurse, supporting its value as a consistent and scalable imaging based tool. AI-based thermography screening shows promise as an effective adjunct to clinical assessments and a foundation for remote monitoring. Further refinement and validation in longitudinal studies may enhance its role in early detection and prevention of diabetic foot complications.