BLOOD GLUCOSE PREDICTION FOR DIABETES THERAPY USING A RECURRENT ARTIFICIAL NEURAL NETWORK William Sandham=, Dimitra Nikoletou==, David Hamilton=, Ken Paterson * , Alan Japp * , Catriona MacGregor= = Dept of Electronic and Electrical Engineering, University of Strathclyde, Glasgow G1 1XW, SCOTLAND Tel: +44 141 552 4400, Fax: +44 141 552 2487 E-Mail: w.sandham@eee.strath.ac.uk, d.hamilton@eee.strath.ac.uk == Bioengineering Unit, Wolfson Centre, 106 Rottenrow, University of Strathclyde, Glasgow G4 0NW, Scotland. Tel: +44 141 552 4400, Fax: +44 141 552 6098 * Diabetes Centre, Royal Infirmary, 84 Castle Street, Glasgow G4 0SF, Scotland. Tel: +44 141 211 4745/4504 E-Mail: ken.mairi@dial.pipex.com ABSTRACT Expert short-term management of diabetes through good glycaemic control, is necessary to delay or even prevent serious degenerative complications developing in the long term, due to consistently high blood glucose levels (BGLs). Good glycaemic control may be achieved by predicting a future BGL based on past BGLs and past and anticipated diet, exercise schedule and insulin regime (the latter for insulin dependent diabetics). This predicted BGL may then be used in a computerised management system to achieve short-term normoglycaemia. This paper investigates the use of a recurrent artificial neural network for predicting BGL, and presents preliminary results for two insulin dependent diabetic females. 1 INTRODUCTION Diabetes mellitus (DM) affects an estimated 3-4% of the world's population (half of whom are undiagnosed), making it one of the major chronic illnesses prevailing today. It is defined as “a syndrome characterised by chronic hyperglycaemia and disturbances of carbohydrate, fat and protein metabolism, associated with absolute or relative deficiencies in insulin secretion and/or insulin action” [1]. The internationally-accepted classification system for DM is Type 1 (insulin-dependent DM or IDDM) and Type 2 (non-insulin dependent DM or NIDDM). Expert short-term management of the condition through good glycaemic control, is necessary to delay or even prevent serious degenerative complications such as retinopathy, neuropathy and nephropathy developing in the long term, due to consistently high blood glucose levels (BGLs) [2]. Potentially life-threatening short-term complications can also occur due to both very low BGLs (hypoglycaemia), and very high BGLs (ketoacidosis). A critical factor which affects BGL in a (Type 1) individual is insulin sensitivity/resistance, which is closely linked to body mass index (weight/height ratio). This tends to be stable within an individual but differs substantially between individuals. External factors which exert a variable influence on a (Type 1) individual's BGL in the short-term include diet, exercise and insulin regime, and these are employed, together with a past record of a patient's BGL, to determine appropriate therapy. Unfortunately, the complex combination of the above factors, together with their short-term dependencies, can lead to prescribed therapies for Type 1 diabetics being sub-optimal. In addition, a considerable proportion of patients are insufficiently knowledgeable regarding their condition, and are therefore unable to alter their short- term therapy confidently, in response to changes in diet or exercise levels, which can again result in sub-optimal therapy. The serious consequences of sub-optimal therapy not only affect a patient's long-term well-being, but lead to a considerable drain on a country's national health resources. A more satisfactory situation would be for the patient to take greater control of his/her short-term therapy, guided by a powerful, yet affordable, computerised management system. Regular attendance at a diabetic clinic would, of course, still be mandatory for longer-term clinical examination, assessment and therapy, in accordance with the St. Vincent Declaration [3]. The longer-term aim of the present study is to investigate the feasibility of using artificial neural networks (ANNs) for educating and advising Type 1 diabetic patients regarding their optimum short-term therapy. Benefits to patients include (a) improved short- term (and hence long-term) therapy, (b) patients encouraged to employ more self-management in the short- term, (c) should help to prevent the occurrence of short- term complications such as hypoglycaemia and ketoacidosis, (d) patient-centred system would be dynamically tailored to the physiology, lifestyle and diabetes condition of the individual, (e) accurate and comprehensive BGL and therapy models, together with appropriate graphical user interface, should aid patient understanding of their condition, and prove useful for teaching purposes in diabetes clinics and medical centres. Finally, a production model with an integral BGL meter,