Abstract submitted for the 59th meeting the European Association for the Study of Diabetes,
EASD, Hamburg, October 2–6, 2023

Evaluation of a Clinical Artificial Intelligence Companion for Therapeutic Decision making in Type 2 Diabetes Management – a Retrospective Proof-of-concept Analysis
Lia Bally1, David Herzig1, Camillo Piazza1,
Lucas Brunschwig2,3, Zeina Gabr2,3, Yaron Dibner2, André Jaun2

1 UDEM, Inselspital, Freiburgstrasse 2, 3010 Bern, Switzerland
2 Metadvice, Route Cantonale 109, 1025 St-Sulpice, Switzerland
3 Ecole Polytechnique Fédérale de Lausanne, 1015 Ecublens, Switzerland

Background and aims. Combining clinical guidelines with continuous learning from real world evidence data has the potential to effectively implement precision medicine by clinicians caring for people with diabetes. Such AI-driven platforms that integrate data from electronic health records are available and are readily extendable to include data from wearable technologies.

Material and methods. Based on the new EASD-ADA guidelines on managing type 2 diabetes and local expert knowledge (including insulin titration protocols), we developed a neuronal network that mimics guideline-conform therapeutic management decisions of diabetes specialists with a 95% accuracy. Following the establishment of data extraction and de-identification pipeline, we evaluated the adherence of clinicians using a retrospective data analysis and assessed to which extent the algorithm may improve glucose control as measured by haemoglobin A1C (HbA1c).

Results. Applying the technology to usual care data from a total of 1,640 patients with diabetes (insulin and non-insulin treated), we find that 60% of a total of 1,548 therapeutic decisions agree with the algorithm. Testing for the impact on glucose control, the weighted average relative reduction in HbA1c over treatment regimens that are in-line with guidelines (–5.62%, N=283) significantly exceeded those that were not (+0.22%, N=362); also, proportion testing for a target relative HbA1c reduction of 5% yields a p-value of 4*10-7 showing clear retrospective outperformance of guidelines therapies over the alternatives, underscoring the potential for optimisation. Using additional input from wearables and transfer-learning from real world evidence outcomes, we will show examples of precision medicine recommendations that are both anchored in medical guidelines and reflect specificities of smaller cohorts justified by glucose and other outcomes (e.g. body weight, lipid control, kidney function).

Conclusion. In this proof-of-concept study, we demonstrated that a neuronal network can mimic clinical guidelines and expert knowledge to improve diabetes control. Prospective studies will provide insights into its usability and therapeutic efficacy in daily clinical practice.

Acknowledgements: Innosuisse – Swiss Innovation Agency (101.082 IP-LS)

Metadvice © 2023