AI tool accurately flags high-risk diabetic patients for neuropathy screening

This pilot study provides essential insights into how AI can augment traditional clinical workflows in diabetes management by identifying patients at risk of complications before symptoms manifest. With a projected global rise in DPN prevalence, non-invasive and scalable methods such as AI-assisted screening could play a pivotal role in preventing severe neuropathic outcomes.

The growing global burden of type 2 diabetes has intensified the search for early detection strategies to prevent debilitating complications such as diabetic polyneuropathy (DPN), a condition that affects nearly half of all diabetic individuals during their lifetime. DPN can progress silently and result in severe consequences such as foot ulcers, infections, and even amputations.

A new study titled "Artificial Intelligence Algorithm to Screen for Diabetic Neuropathy: A Pilot Study", published in Biomedicines, evaluates the performance of an AI-based clinical prediction tool to stratify patients at risk of DPN and guide timely screening using biothesiometry testing.

How does the AI model predict diabetic neuropathy risk?

The study was conducted at the Diabetes Operating Unit of the ULSS 6 in Padua, Italy, involving 201 asymptomatic patients with type 2 diabetes aged between 18 and 80 years. Researchers used the MetaClinic Prediction Algorithm, an AI tool developed by METEDA S.r.l., to assign each patient a risk level of developing DPN within a two-year horizon. The AI algorithm leverages demographic, clinical, and biochemical data such as blood glucose, cholesterol levels, renal function markers, blood pressure, and HbA1c to generate risk scores across four categories: low, moderate, high, and very high.

These AI-generated risk scores were then compared to results from a biothesiometer, a non-invasive diagnostic device that measures the vibratory perception threshold (VPT) in patients’ feet - a gold standard method for diagnosing large-fiber peripheral neuropathy. A VPT score ≥ 25 V was considered indicative of DPN.

Results showed that 107 patients were classified as low risk, 39 as moderate, 29 as high, and 26 as very high. In total, 63 patients (31.34%) had a VPT ≥ 25 V, confirming the presence of DPN. The remaining 138 patients had normal VPT scores below 25 V. The AI model showed a 65% overall agreement with the biothesiometer, with Cohen’s κ at 0.162 and Gwet’s AC1 coefficient at 0.405, indicating slight to fair agreement beyond chance.

How well did the AI model align with biothesiometer testing?

While the AI system demonstrated potential in risk stratification, its concordance with objective biothesiometry testing was modest. Among the 63 patients with confirmed DPN, 26 had been classified as low risk by the AI, suggesting a notable rate of false negatives. Conversely, some individuals flagged as high or very high risk by the AI showed no abnormal VPT, indicating potential false positives. These discrepancies could stem from the AI's reliance on general metabolic markers rather than direct neural indicators.

The study found that patients with DPN were significantly older (average 72.7 years vs. 66.6 years) and had a longer disease duration (14.4 vs. 11 years) compared to those without neuropathy. These two variables, age and diabetes duration, were not direct inputs in the AI model used but proved to be strong differentiators, reinforcing their role as important risk factors for DPN development. Additionally, diastolic blood pressure was slightly lower in patients with DPN (77.2 mmHg vs. 80.5 mmHg), though no significant differences emerged in lipid profiles, glucose levels, or kidney function indicators.

The moderate concordance between the AI predictions and biothesiometer results highlights both the potential and limitations of current AI tools in diabetic care. The authors suggest that the AI model’s predictive power could be enhanced by incorporating additional risk markers more specific to neuropathic changes, including small-fiber nerve function, foot temperature variation, or glycemic variability.

Despite these limitations, the AI tool serves a valuable function as a triage mechanism. Patients classified as high or very high risk could be prioritized for further screening with biothesiometry or more advanced neurophysiological evaluations, improving resource allocation and early intervention.

What are the broader implications for clinical practice and future research?

This pilot study provides essential insights into how AI can augment traditional clinical workflows in diabetes management by identifying patients at risk of complications before symptoms manifest. With a projected global rise in DPN prevalence, non-invasive and scalable methods such as AI-assisted screening could play a pivotal role in preventing severe neuropathic outcomes.

Crucially, the AI algorithm used in this study was trained on a massive dataset of over 147,000 patients across 23 diabetes centers in Italy and was externally validated using data from five independent centers. However, the specific validation in this pilot involved a new patient population, strengthening the generalizability of findings while exposing areas needing refinement.

Compared to other emerging methods, such as AI models based on corneal confocal microscopy, magnetic resonance neurography, or toe photoplethysmography, the MetaClinic algorithm offers superior scalability because it uses routinely collected clinical data. However, the authors acknowledge that a lack of specificity and limited incorporation of direct neurological indicators remain challenges.

Future iterations of the AI model could integrate disease duration, patient age, and sensory testing results to improve precision. The study also highlights the need for broader multi-center trials involving symptomatic and diverse patient populations to assess real-world effectiveness and identify subgroups who benefit most from AI-guided triage.

From a practical standpoint, embedding AI tools into electronic health records, educating clinicians about interpreting AI outputs, and ensuring alignment with screening guidelines will be critical for adoption. The biothesiometer, a portable, low-cost instrument, presents an ideal partner for AI tools in resource-limited settings, especially as global diabetes rates surge.

Ethical considerations, including data privacy, fairness, and explainability of AI predictions, must also be addressed. Ensuring that AI complements, rather than replaces, clinician judgment is essential to maintain patient trust and clinical accountability.