Validation of the effectiveness of a digital integrated healthcare platform utilizing an AI-based dietary management solution and a real-time continuous glucose monitoring system for diabetes management: a randomized controlled trial

The researchers planned to compare three groups:

**Group A (Control):** Routine diabetes care — standard hospital visits every 3 months, no digital platform.

**Group B (Platform only):** Used a digital integrated healthcare platform (app called "AutoChek Care") that automatically collected blood glucose, weight, body fat, blood pressure, and step count via Bluetooth devices. They also used an AI-based dietary management app called "FoodLens" that identifies foods from a single photo and estimates calories and nutrients. No feedback from medical staff.

**Group C (Platform + CGM + Feedback):** Same platform as Group B, plus a continuous glucose monitoring system (Dexcom G5) worn for 1 week every 3 months, and weekly text message feedback from a clinical dietitian about glucose levels, weight control, diet, and exercise.

The **primary outcome** was change in HbA1c from baseline to 6 months (24 weeks). **Secondary outcomes** included HbA1c change at 12 months, body weight, body fat mass, exercise amount, dietary information, lipid profile, CGM metrics (time in range, etc.), Diabetes Treatment Satisfaction Questionnaire (DTSQ) scores, and number of hypoglycemic events or complications.

The study planned to recruit **patients with type 2 diabetes mellitus (T2DM)** from three university-affiliated hospitals in Seoul, South Korea (Samsung Medical Center, Asan Medical Center, Kyung Hee University Hospital).

**Inclusion criteria:**

Age 19–69 years

Diagnosed with T2DM

HbA1c between 53–69 mmol/mol (7.0–8.5%) — moderately elevated but not severely uncontrolled

BMI ≥23 kg/m² (overweight threshold for Asian populations)

Either no hypoglycemic medication in the prior 4 weeks, or stable dose of one or more oral hypoglycemic agents for ≥12 weeks

Willing to use the digital platform and sign consent

**Exclusion criteria:**

Type 1 diabetes or gestational diabetes

Using insulin or GLP-1 agonist injections

Uncontrolled chronic liver disease, acute kidney injury, psychiatric disorders (schizophrenia, depression, bipolar disorder)

Using anti-obesity medication, corticosteroids, or immunosuppressants

Alcoholism or drug addiction within 3 months

Pregnancy, lactation, or planning pregnancy

The target sample size was not explicitly stated in the protocol excerpt, but the trial registration (NCT04161170) typically specifies around 150–200 participants (50–67 per group) to detect a clinically meaningful HbA1c difference.

**HbA1c:** Blood test at baseline and every 12 weeks (weeks 12, 24, 36, 48). This is the gold-standard measure of average blood glucose over ~3 months.

**Body weight and body fat mass:** Bluetooth-connected scale with bioelectrical impedance analysis, automatically transmitted to the app.

**Blood pressure:** Bluetooth-connected sphygmomanometer, automatically transmitted.

**Physical activity:** Watch-type pedometer measuring steps, calories burned, and walking distance, automatically transmitted.

**Dietary intake:** AI-based "FoodLens" app — users photograph their meal before eating; the app identifies multiple foods in one photo and estimates nutritional composition and calories. Reported recognition accuracy: 86.6% for food classification.

**Continuous glucose monitoring (Group C only):** Dexcom G5 — sensor measures interstitial fluid glucose every 5 minutes, transmits to smartphone via Bluetooth 4.0. Provides real-time glucose values, trends, and alerts for hypo/hyperglycemia. Worn for 1 week every 3 months.

**Diabetes Treatment Satisfaction Questionnaire (DTSQ):** Validated questionnaire measuring satisfaction with diabetes treatment. Administered at every visit.

**Lipid profile and other lab tests:** Blood and urine tests at each visit.

**Hypoglycemic events:** Recorded throughout the study.

**Study design:** This is a **48-week, open-label, randomized, multicenter, parallel-group trial**. It is a **protocol paper** — meaning the study was designed and registered, but results were not yet reported in this publication.

**Randomization:** Participants were to be randomly assigned to three groups in a 1:1:1 ratio. The randomization method was not detailed in the protocol excerpt, but multicenter trials typically use computer-generated random sequences with stratification by site.

**Blinding:** This is an **open-label** study — no blinding. Participants, medical staff, and outcome assessors all know which group each participant is in. This is a major limitation because:

Participants in Group C know they are getting extra attention and a CGM device, which could create a placebo effect (they may change behavior simply because they feel monitored).

Medical staff providing feedback in Group C also know the treatment assignment, which could bias their interactions.

Outcome assessment (HbA1c) is objective, so it is less susceptible to bias than subjective outcomes like satisfaction scores.

**Duration:** 48 weeks total (12 months). The primary endpoint is at 24 weeks (6 months). Follow-up continues to 48 weeks to assess durability of effects.

**Washout periods:** Not applicable — this is a parallel-group design, not a crossover. Participants remain in their assigned group for the full 48 weeks.

**Statistical approach:** The protocol mentions comparing change in HbA1c from baseline to 6 months between groups, likely using analysis of covariance (ANCOVA) adjusting for baseline HbA1c and other covariates. Intention-to-treat analysis is standard for RCTs.

**What this design can prove:**

If successful, it can show that the digital platform with CGM and feedback causes greater HbA1c reduction than routine care (causality, due to randomization).

The three-group design allows separating the effect of the platform alone (Group B vs. A) from the effect of platform plus CGM plus feedback (Group C vs. A), and also comparing Group C vs. B to isolate the added value of CGM and medical staff feedback.

**What this design cannot prove:**

It cannot prove which specific component of Group C's intervention is responsible for any observed effect — is it the CGM? The weekly feedback? The AI diet app? The combination? The design only compares the whole package vs. routine care.

It cannot prove long-term effects beyond 12 months.

It cannot prove effectiveness in patients with more severe diabetes (HbA1c >8.5%), those on insulin, or those with normal BMI.

Open-label design means any subjective outcomes (satisfaction, adherence) could be biased.

**Major methodological weaknesses:**

**No blinding** — the most significant weakness. Placebo effects are well-documented in diabetes interventions.

**Protocol only** — no results are reported. This paper describes what they *plan* to do, not what they *found*.

**Industry involvement** — the platform (AutoChek Care) is developed by Aprillis Co., Ltd., and the AI diet app (FoodLens) by DoingLab Co., Ltd. Potential conflict of interest.

**No sham control** — Group A gets no device or app, so any effect could be due to the attention and novelty of using technology, not the specific features.

**Short CGM wear time** — only 1 week every 3 months may miss important glucose variability patterns.

**This is a study protocol — no results are reported.** The paper describes the planned methodology, not the outcomes. Key planned analyses include:

**Primary endpoint:** Change in HbA1c from baseline to 24 weeks. The researchers hypothesized that Group C would show greater reduction than Group A, and Group B would show greater reduction than Group A.

**Secondary endpoints:** Changes in HbA1c at 48 weeks; changes in body weight, body fat mass, exercise amount, dietary intake, lipid profile; CGM metrics (time in range, time above range, time below range, glucose variability); DTSQ scores; number of hypoglycemic events.

No effect sizes, p-values, or confidence intervals are available because the study had not yet been completed at the time of publication.

Not applicable — no results are reported. However, the study was designed to detect a clinically meaningful difference in HbA1c. Typically, diabetes trials aim for a 0.3–0.5% (3–5 mmol/mol) reduction in HbA1c as clinically significant. The inclusion criteria (HbA1c 7.0–8.5%) suggest the researchers expected the intervention to lower HbA1c by at least 0.5% compared to routine care.

**Acknowledged by authors:**

The protocol mentions that the study is open-label, which they acknowledge as a limitation.

They note that the AI food recognition accuracy is 86.6%, meaning ~13% of foods may be misidentified.

**Critical reader observations:**

**No results** — this is a protocol, so no conclusions can be drawn about effectiveness.

**Open-label design** — the most serious limitation. Participants who know they are getting a high-tech intervention with medical staff attention may change their behavior (Hawthorne effect), and medical staff may unconsciously provide better care to the intervention group.

**Industry funding and involvement** — the platform and AI app are commercial products from companies that may benefit from positive results. This creates potential for publication bias.

**Short CGM wear** — 1 week every 3 months may not capture representative glucose patterns, especially if participants change their behavior during the monitoring week.

**Population limits** — only Asian patients with overweight (BMI ≥23), moderate hyperglycemia (HbA1c 7.0–8.5%), and no insulin use. Results may not generalize to other ethnicities, normal-weight individuals, those with more severe diabetes, or those on insulin.

**No sham control** — Group A receives no device or app, so any benefit could be due to the novelty of using technology rather than the specific features.

**Multiple comparisons** — many secondary outcomes increase the risk of false-positive findings.

**Attrition risk** — 48 weeks is a long study; dropouts may differ between groups and bias results.

**No objective adherence measure** — while devices transmit data automatically, there is no way to verify that participants actually used the AI diet app for every meal or wore the pedometer consistently.

For someone running their own n=1 experiment to improve blood sugar control using similar tools:

**What to test:**

Test the combination of: (1) photographing every meal before eating using a food recognition app (like FoodLens, MyFitnessPal with photo feature, or similar), (2) wearing a continuous glucose monitor (CGM) for at least 2 weeks continuously (not just 1 week per quarter), and (3) reviewing your own glucose data daily and adjusting meals based on postprandial spikes.

If you want to isolate the effect of feedback, try 4 weeks of self-monitoring alone, then add weekly review with a coach or accountability partner.

**Minimum meaningful duration:**

At least 4–6 weeks for CGM data to capture enough meals and activities to see patterns.

At least 12 weeks (3 months) to see a meaningful change in HbA1c, since HbA1c reflects average glucose over ~3 months.

For a self-experiment, aim for 12 weeks minimum, with daily CGM wear throughout (not just 1 week).

**What to measure (specific metrics):**

**Primary:** Fasting glucose (daily) and estimated HbA1c from CGM (or lab HbA1c at start and end).

**Secondary:** Time in range (TIR) — percentage of readings between 70–180 mg/dL (3.9–10.0 mmol/L). Aim for >70% TIR.

**Secondary:** Time above range (>180 mg/dL) — aim for <25%.

**Secondary:** Time below range (<70 mg/dL) — aim for <4%.

**Secondary:** Glucose variability — standard deviation or coefficient of variation of glucose readings. Lower is better.

**Secondary:** Postprandial glucose spikes — peak glucose 1–2 hours after meals. Aim for peak <180 mg/dL.

**Secondary:** Body weight (daily, same time each morning after voiding).

**Secondary:** Number of hypoglycemic events (glucose <70 mg/dL with symptoms).

**Key confounds to control for:**

**Medication changes** — keep oral diabetes medications at the same dose throughout the experiment. If your doctor changes your meds, restart the experiment.

**Exercise** — log type, duration, and intensity. Exercise improves insulin sensitivity for 24–48 hours. Try to keep exercise consistent week-to-week, or at least record it so you can account for it.

**Sleep** — poor sleep increases insulin resistance. Log sleep duration and quality. Aim for 7–9 hours per night.

**Stress** — acute and chronic stress raise blood glucose via cortisol. Log daily stress level (1–10 scale).

**Illness** — infections, colds, and inflammation raise blood glucose. Exclude sick days from analysis.

**Alcohol** — alcohol can cause delayed hypoglycemia (especially at night). Log alcohol consumption.

**Menstrual cycle** — insulin sensitivity varies across the cycle. If you menstruate, track cycle phase.

**Meal timing and composition** — not just what you eat, but when. Log meal times and macronutrient composition (carbs, protein, fat, fiber).

**What a positive result would look like:**

HbA1c drops by ≥0.3% (3 mmol/mol) from baseline to 12 weeks.

Time in range increases by ≥10 percentage points (e.g., from 60% to 70%).

Average fasting glucose drops by ≥10 mg/dL (0.6 mmol/L).

Postprandial glucose spikes are reduced by ≥20 mg/dL (1.1 mmol/L) at 1 hour.

Glucose variability (standard deviation) decreases by ≥10 mg/dL.

You experience fewer hypoglycemic events (if you were having them).

Body weight decreases by ≥2 kg (if weight loss is a goal).

**Important caveat for n=1 experiments:** You cannot randomize or blind yourself, so you cannot prove causation. The best you can do is establish a baseline (2–4 weeks of measurement without the intervention), then introduce the intervention and measure for 12+ weeks. If you see a clear, sustained change in the expected direction, that is suggestive but not definitive. Consider repeating the experiment (ABAB design: baseline → intervention → return to baseline → intervention again) to strengthen the evidence.