Cost-effectiveness and clinical outcomes of artificial intelligence-enhanced screening for diabetic foot ulcers: A simulation study

Dear Editor,

Diabetic foot ulcers (DFUs) are a serious complication of diabetes mellitus, with a lifetime risk estimated to be between 19% and 34%.¹ Without timely prevention and management, DFUs can lead to lower extremity amputations (LEAs) and premature death.^2,3 DFUs also impose significant healthcare and societal costs, especially in Southeast Asia.^4,5 Regular foot screenings are essential for preventing these complications.

Countries in Southeast Asia generally follow the guidelines set by the 2023 International Working Group on the Diabetic Foot,⁶ which recommend foot screening frequency based on their risk stratification system: annually for very low risk, once every 6–12 months for low risk, once every 3–6 months for moderate risk and once every 1–3 months for high risk. However, adherence to these guidelines reveals significant challenges and variability in practice due to barriers such as resource limitations and staff shortages. In Singapore, the Agency for Care Effectiveness recommends differentiated screening intervals: annually for low-risk patients, every 6 months for moderate-risk patients and every 3–4 months for high-risk patients.⁷ Low-risk patients are defined as those without any of the following risk factors: calluses, deformities, peripheral artery disease, neuropathy, previous foot ulcers or amputations. Evidence⁸ suggests that the risk of ulceration in low-risk patients remains stable over time. We estimated that 60–80% patients with diabetes can be classified as low risk, and their annual DFU incidence risk is lower than 1% based on data from our cluster’s Chronic Disease Management Datamart (CDMD),⁹ raising concerns about overscreening and inefficient resource allocation.

Artificial intelligence (AI) offers a potential solution by enabling personalised risk-tailored screening.¹⁰ This study evaluated the cost-effectiveness and clinical outcomes of AI-enhanced screening versus routine annual screening for low-risk patients. Using an XGBoost-based predictive model, the AI-enhanced screening approach involves screening based on individual risk: patients flagged as positive are screened annually, while negative low-risk patients are screened every 3 years. The predictive model achieved an area under the receiver operating characteristics curve of 0.81, with a sensitivity of 0.7 and a specificity of 0.8. A cohort of 500,000 low-risk patients with diabetes in Singapore, with an average age of 50 years, was simulated over a lifetime to assess the long-term impact of the AI-enhanced screening.

A Markov state-transition model was constructed to simulate disease progression through 5 health states: diabetes, DFU, minor LEA, major LEA and death. Model parameters, including both base values and distributions of disease transition probabilities and costs (Table 1), were derived from CDMD⁹ and validated against international sources.^11,12 Direct medical cost for foot screening, examination, consultation, DFU treatment and management, LEA procedures, and related services were included in the cost analysis. Effectiveness was measured in quality-adjusted life years (QALYs), based on EuroQoL-5 Dimension data from a Singapore study.¹³ A 3% discount rate was applied for both costs and effectiveness. The incremental cost-effectiveness ratio (ICER) was then assessed to determine the most cost-effective screening strategy using gross domestic product (GDP) per capita as the willingness-to-pay (WTP) threshold as commonly used in Singapore. Other outcomes measured in the simulation included the number of patients screened, DFU detections, minor and major LEAs performed, and deaths. Monte Carlo micro-simulations with 1000 samples, randomly drawn from parameter distributions, were performed for probability sensitivity analysis (PSA) to account for uncertainties in model parameters. Results were summarised in cost-effectiveness acceptability curves (CEAC) and expected loss curves (ELC) to determine the optimal screening strategy under uncertainty.¹⁴ The CEAC shows the chance of each strategy being the most cost-effective and picks the one with the highest chance. The ELC shows the potential cost of choosing the wrong strategy and selects the one with the lowest risk of loss.

Table 1. Model parameters and their values^a used in the study.

AI-enhanced screening resulted in an average lifelong cost of SGD54,272 per patient, compared to SGD55,587 for annual screening, saving SGD1315 per patient. The lifelong effectiveness for AI-enhanced screening was 23.150 QALYs per patient, which is slightly lower than 23.154 QALYs for annual screening with a minimal difference of 0.004 QALYs per patient. AI-enhanced screening required 4,372,523 DFU screenings, compared to 11,178,861 screenings under annual screening. LEA rates were similar between strategies, with 52,720 minor and 68,646 major LEAs for AI-enhanced screening versus 52,689 minor and 68,710 major LEAs for routine screening. AI screening led to 26 additional deaths. ICER was SGD292,181 per QALY gained, well below Singapore’s GDP per capita in 2023 (SGD110,000), indicating that AI-enhanced screening was more cost-effective than annual screening.

PSA showed that AI-enhanced screening had an average cost of SGD52,178 and effectiveness of 22.866 QALYs, compared to SGD53,993 and 22.876 QALYs for routine screening. ICER from PSA was SGD174,572 (standard deviation: SGD13,296) per QALY gained. Both CEAC and ELC indicated that AI-enhanced screening was optimal when WTP was below SGD180,000 per QALY.

AI-enhanced DFU screening optimises healthcare resource allocation by tailoring screening intervals based on individual risk, leading to long-term cost savings. This aligns with government initiatives to reduce healthcare expenditure while maintaining high standards of care, benefiting both patients and healthcare systems. In resource-constrained settings, this approach could alleviate healthcare burdens and enable better resource allocation. The findings are relevant not only for Singapore but also for other countries in Southeast Asia and globally, where diabetes incidence is high and healthcare resources are limited. Malaysia, Indonesia and Thailand could benefit from similar AI-driven strategies, adapted to local healthcare systems. The predictors used in our AI model are routinely collected, facilitating easy adoption and adaptation across different healthcare settings. This flexibility allows the model to be tested and validated in diverse regions, ensuring it can be adjusted for local infrastructure, patient demographics and resource availability. Such adaptability ensures broader applicability and smoother integration into various health systems.

In conclusion, this study demonstrates that AI-enhanced DFU screening can offer significant cost savings while maintaining high standards of care. By tailoring screening intervals based on individual risk, healthcare systems can improve resource efficiency while maintaining high standards of care. Future research should focus on validating these findings in diverse populations and addressing ethical concerns to ensure that AI-driven healthcare innovations are accessible to all.

REFERENCES

1. Armstrong DG, Boulton AJM, Bus SA. Diabetic Foot Ulcers and Their Recurrence. N Engl J Med 2017;376:2367-75.
2. Crocker RM, Palmer KNB, Marrero DG, et al. Patient perspectives on the physical, psycho-social, and financial impacts of diabetic foot ulceration and amputation. J Diabetes Complications 2021;35:107960.
3. Sekhar MS, Thomas RR, Unnikrishnan MK, et al. Impact of diabetic foot ulcer on health-related quality of life: A cross-sectional study. Semin Vasc Surg 2015;28:165-71.
4. Lo ZJ, Surendra NK, Saxena A, et al. Clinical and economic burden of diabetic foot ulcers: A 5‐year longitudinal multi‐ethnic cohort study from the tropics. Int Wound J 2021;18:375-86.
5. Ang GY, Yap CW, Saxena N. Effectiveness of Diabetes Foot Screening in Primary Care in Preventing Lower Extremity Amputations. Ann Acad Med Singap 2017;46:417-23.
6. Schaper NC, van Netten JJ, Apelqvist J, et al. Practical guidelines on the prevention and management of diabetes-related foot disease: IWGDF 2023 update. The International Working Group on the Diabetic Foot 2023. https://iwgdfguidelines.org/wp-content/uploads/2023/07/IWGDF-Guidelines-2023.pdf. Accessed 17 September 2024.
7. Agency for Care Effectiveness. Foot assessment in patients with diabetes mellitus, 6 June 2019. Updated 8 August 2024. https://www.ace-hta.gov.sg/docs/default-source/acgs/foot-assessment-in-patients-with-diabetes-mellitus-(aug-2024).pdf?sfvrsn=7ee3b22c_5. Accessed 17 September 2024.
8. Crawford F, Chappell FM, Lewsey J, et al. Risk assessments and structured care interventions for prevention of foot ulceration in diabetes: development and validation of a prognostic model. Health Technol Assess 2020;24:1-198.
9. Gunapal PPG, Kannapiran P, Teow KL, et al. Setting up a regional health system database for seamless population health management in Singapore. Proceedings of Singapore Healthcare 2016;25:27-34.
10. Tan EC. Artificial Intelligence and Medical Innovation. Ann Acad Med Singap 2020;49:252-5.
11. Cheng Q, Lazzarini PA, Gibb M, et al. A cost‐effectiveness analysis of optimal care for diabetic foot ulcers in Australia. Int Wound J 2016;14:616-28.
12. Boodoo C, Perry JA, Leung G, et al. Cost-effectiveness of telemonitoring screening for diabetic foot ulcer: a mathematical model. CMAJ Open 2018;6:E486-94.
13. Lo ZJ, Tan E, Chandrasekar S, et al. Diabetic foot in primary and tertiary (DEFINITE) Care: A health services innovation in coordination of diabetic foot ulcer (DFU) Care within a healthcare cluster ‐ 18‐month results from an observational population health cohort study. Int Wound J 2023;20:1609-21.
14. Wolff HB, Qendri V, Kunst N, et al. Methods for Communicating the Impact of Parameter Uncertainty in a Multiple-Strategies Cost-Effectiveness Comparison. Med Decis Making 2022;42:956-68.