Study Protocol for the Effects of Artificial Intelligence (AI)-Supported Automated Nutritional Intervention on Glycemic Control in Patients with Type 2 Diabetes Mellitus

This is a multicenter, unblinded, parallel, randomized controlled study aimed at comparing the efficacy of AI-supported automated nutrition therapy with that of conventional human nutrition therapy in patients with type 2 diabetes. The study protocol was approved by the Institutional Review Board of Kanazawa University on April 13, 2018 (IRB no. 2623-3); all recruited patients are required to provide written informed consent. This trial is registered in the University Hospital Medical Information Network (UMIN-CTR; https://​www.​umin.​ac.​jp/​ctr; UMIN ID: UMIN000032231), and is being conducted in accordance with the Declaration of Helsinki of 1964, as revised in 2013. Patient recruitment commenced on April 13, 2018 and is scheduled to continue through to April 13, 2020.

Participants are recruited from patients attending the outpatient clinics of the Department of Diabetes and Endocrinology at Kanazawa University Hospital, Hokuriku Central Hospital, and Houju Memorial Hospital. Those patients who are obese or overweight and have type 2 diabetes mainly controlled with diet are eligible.

Inclusion criteria are (1) age ≥ 20 years; (2) body mass index (BMI) cutoff for Asians of > 23 kg/m², which is the World Health Organization cutoff for Asians requiring public health action [16]; (3) physician diagnosis of type 2 diabetes confirmed by clinical data (e.g., documentation of fasting plasma glucose [FPG] of ≥ 126 mg/dL, a positive oral glucose tolerance test of > 200 mg/dL, or HbA1c ≥ 6.5%); (4) owning and using a mobile phone in their daily lives; and (5) able to understand study procedures and provide written informed consent.

Exclusion criteria include (1) use of insulin and/or glucagon-like peptide-1 within the past 6 months; (2) use of agents causing weight loss; (3) HbA1c ≥ 7.5%; (4) renal dysfunction (e.g., urinary protein > 1 g /g creatinine and/or an estimated glomerular filtration rate of < 45 mL/min/1.73 m²); (5) uncontrolled thyroid disease; (6) uncontrolled hypertension (diastolic blood pressure > 100 mm Hg and/or systolic blood pressure > 160 mm Hg) and/or secondary hypertension; (7) oral or injected steroid use within the past 3 months; (8) pregnancy or planned pregnancy during the study period; (9) history of myocardial infarction, severe uncontrolled heart failure, unstable angina, transient ischemic attack, or stroke in the past 6 months; (10) receiving treatments for malignancies; (11) diagnosed eating disorder; and (12) medical or psychological condition(s) which would compromise the ability to meet protocol requirements in the opinion of the investigator.

Patients are allocated randomly to either AI- and mobile-supported nutrition therapy or in-person nutrition therapy by human dieticians for 1 year. The principle of nutrition therapy in both interventions is total energy restriction, based on the recommendations of Japan Diabetes Society (JDS), Japan Atherosclerosis Society (JAS), and Japan Society for the Study of Obesity (JASSO); the ideal body weight (IBW) is calculated as IBW = height (m)² × 22 (kg/m²) [17, 18], and the total dietary energy (kcal/day) is estimated as 25–35 kcal/kg IBW/day. Physicians determine the total dietary caloric intake according to the daily activity of each patient [17, 18]: 25–30 kcal/kg IBW for subjects with low-level activity, 30–35 kcal/kg IBW for subjects with usual-level activity, and ≥ 35 kcal/kg IBW for subjects with heavy-level activity. The recommended macronutrient distribution is set to be 50–60% from carbohydrate sources, ≤ 20% from protein sources, and 20–30% from fat sources [17].

The mobile phone app used for this study is called “Asken” and is one of the most popular Japanese apps for behavior change among individuals aspiring to lose weight. The criteria of its nutritional feedback system has been specially modified for this study so that it follows the recommendations set by the JDS. One of the main features of this app is the availability of various food-logging interfaces, most notably an AI-powered photo analysis system linked with its extensive database of over 100,000 menus both on the market and home-made. Once a photo of a hamburger combo is taken, for example, the system instantly identifies the frame of each item as well as its menu and serving amount (Fig. 1). After confirmation by the patient, during which process the patient can re-select the correct type or brand of hamburger from 82 types/brands in the database, the app calculates the energy and nutrient intakes based on the Standard Tables of Food Composition in Japan 2015, seventh revised edition [19]. The validity and accuracy of Asken’s estimated total energy intake and macronutrients, such as carbohydrates, fat, and protein, have been demonstrated in an earlier study [15]. Following each food logging, Asken eventually shows how the user has satisfied his/her recommended range of daily intake of each nutrient with a graph (Fig. 2) and delivers an individualized message by AI based on the Dietary Reference Intakes for Japanese 2015 [20]. Because this whole process from menu identification with photo analysis to generation of dietary feedbacks is fully unmanned and automated, the patients can learn their detailed nutritional conditions anywhere and anytime. These dietary messages and feedbacks are also delivered by a female Japanese character “Miki”; there are more than 200,000 different patterns for her wording and facial expressions. “Miki” always remains compassionate and empathetic to the patients even when they eat a lot or skip a recording of food intake (Fig. 3).

The patients also have access to peer support from other Asken users through the in-app diary function. Weights and steps can be automatically recorded in the app by its data-sharing function with other apps, such as GoogleFit and Apple Healthcare. Patients are instructed to log onto the Asken app and input meals and snacks at least once a week.

In Japan, nutrition therapy managed by human dietitians is covered by health insurance programs for patients with diabetes. The principle of nutrition therapy of a comparison group is the same as that of the experimental group, namely, total caloric restriction with macronutrient distribution as recommended by the JDS [17]. At baseline, dietitians estimate the current energy and nutrient intakes on a face-to-face interview. Participants and dietitians work together to design feasible strategies to improve dietary habits. Patients are provided with printed materials for describing their plans and recording their meals. After the first interview at baseline, human dietitians continue to evaluate the achievement status of the initial plans based on the paper records and to provide individual feedbacks once or more times within 3 months. Patients have a one-to-one meeting for ≥ 20 mi with a human dietitian during each visit.

The primary outcome of this study is the change in HbA1c levels during 1 year of intervention, with the outcomes to be compared between the two groups. HbA1c is an accepted diagnostic parameter by which to monitor long-term glycemic control [21]. Secondary outcomes are absolute changes in (1) FPG; (2) plasma lipid profile, total cholesterol, triglycerides, high-density lipoproteins, and low-density lipoproteins; (3) body weight, BMI, and waist circumference; (4) systolic and diastolic blood pressure; and (5) urinary albumin excretion. (6) Recorded data on the Asken app (frequencies of log-ons, weight, daily intakes of total energy and nutrients) and (7) costs of the interventions in economic terms are also evaluated.

Venous blood samples are drawn from the antecubital vein after an overnight fast by a trained nurse and transported to the central laboratory of each hospital for analysis. HbA1c is measured using a high-performance liquid chromatography method. Lipids are measured using enzymatic analytical chemistry methods, and plasma glucose is assessed using the glucose oxidase method. Resting blood pressure is measured in the sitting position with an automatic device after ≥ 5 min of rest. Measurements of BMI and waist circumference are conducted according to published methods [22].

In principle, treatments for diabetes, hypertension, and dyslipidemia are not changed during the study period in order to avoid effects on the efficacy assessments of this study. If it is inevitable to add a new antidiabetic drug to the treatment regimen, a full description of the change and the reason and timing are described in the case report form (CRF). If insulin therapy is started, the trial will be discontinued at that time.

User privacy is preserved from two directions, the app and humans. First, the app is equipped with security policy and measures according to the JIPDEC PrivacyMark® system, which is a widespread certification for companies managing personal information in Japan. The app’s operating company, WIT Co. Ltd., has a secured data server in Japan to deal with all of the user information safely and appropriately while disclosing in the privacy policy what kinds of data are collected for the service and how and by whom the data can be used. In addition, the app data are stored and managed anonymously, kept re-identifiable only by the limited number of authorized researchers. Data managers store the file for such re-identification separately from the study data itself with password-protected access systems in an area with a limited access.

Clinical data of the participants will be extracted from electronic medical records and anonymously provided on an electronic CRF (eCRF) to the data manager. All entries in the CRFs should be backed up with the relevant source data, including informed consent forms, medical history, laboratory data, and recorded data on the app.

Monitoring of this research will be conducted by Innovative Clinical Research Center, Kanazawa University (iCREK). Monitors will ensure that the investigational team complies with the study protocol and that data and adverse events (AEs) are accurately recorded in the CRFs. Due to the low risk character of the intervention in this study, only spontaneously reported AEs will be reported, and the investigators will immediately take appropriate measures when AEs occur.

The schedule for enrollment, interventions, and assessments are shown in Fig. 4. Participants in either intervention group are seen at least once every 3 months for the collection of study outcome data. Data collection is performed during the patient’s routine visits for diabetes checks.

The primary endpoint of this study is the change of HbA1c from baseline to endpoint (12 months of intervention). Based on the decrease in HbA1c in prior studies involving mobile-supported interventions [23], we expect that the mean change in HbA1c level will be 0.3% with a standard deviation of 0.3%. At a significance level of 0.05 and a detection power of 80%, the number of cases per group necessary to reach significant difference in change in HbA1c between two groups is calculated to be 74 cases. Assuming a 20% dropout rate, the target number of cases for each group is 50 cases, with a total of 100 participants in this study.

We use a web-based computerized blocked randomization with blocks of six to assign participants to either two modes of nutrition therapy. Block randomization is performed in each center separately. Randomization is conducted centrally and independently of investigators, and participants in each group are treated equally by their physicians. Although blinding of medical staff and participants to the study allocation is impossible, staff who are responsible for data collection are blinded.

The primary objective of this study is to confirm that AI-supported, fully-automated nutritional intervention is not be inferior to in-person nutrition therapy by human dieticians with regard to HbA1c change during a 1-year period. Non-inferiority is considered confirmed with the use of a prespecified non-inferiority margin of 0.2%. This is established if the lower limit of the 95% confidence interval of the difference in HbA1c change between the two groups (AI-supported group minus human-supported group) is 0.2%. An independent t test is used to compare the two groups for the change in HbA1c, which is calculated by subtracting the value of the baseline from the value after 1 year of intervention. The changes in secondary continuous outcomes of this study are also analyzed using an independent t test. To account for non-dietary factors, such as the addition of new drugs, compliance to medications, and exercise habits, multiple regression analyses for changes in continuous outcomes are performed using these non-dietary factors as covariates.

Participants who discontinue using Asken or to seek advice according to the planned consultations with dieticians will continue to have data collected from their routine diabetes clinic visits, unless they specifically withdraw consent for this study.