Study analysis · The American journal of clinical nutrition · 2023

Your fancy blood sugar app is useless — here's why.

A cheap low-fat diet and a high-tech app improved blood sugar the same tiny amount — the app didn’t help.

Reading level
High certainty
Level 1b · Individual RCT

Overview

What the study found

The study in plain English — the bottom line, every takeaway we extracted, and what to do with them.

In simple terms

This study compared two diets: one that gave everyone the same advice, and one that used a computer app to give personalized food tips. It found that both diets helped lower blood sugar about the same amount. So, the fancy app didn’t make a big difference. That means we can’t say the app causes better blood sugar control — it just didn’t help more than regular advice.

What’s the bottom line?

Scientists tested if a fancy app that tells you which foods spike your blood sugar works better than just eating fewer calories on a low-fat plan.

How strong is this study?

This study did a good job by randomly assigning people to groups so it’s fair, but it had some problems: people knew which diet they were on, many didn’t use the app much, and not everyone stayed in the study. Because of that, we can trust that the diets didn’t differ much, but we should be careful saying the app definitely doesn’t work for anyone — maybe it just didn’t work well here.

Reporting

75 / 100

  • COI disclosure+40/40
  • Data availability+35/35
  • Code availabilitycode not shared
Methodology

70 / 100

  • Randomization+20/20
  • Blindingnot blinded
  • Control group+15/15
  • Sample size (n=156)+10.8/20
  • Follow-up+10/10
Publication

100 / 100

Statistical

100 / 100

  • P-values+15/15
  • Effect size+20/20
  • Confidence intervals+15/15
  • Pre-registration+15/15

Each component is scored out of 100 and then capped by the study design — a case series cannot reach the ceiling a randomised trial can, however well it is reported.

Where it sits

RCT reviews

Max 100

Randomized Trials

Max 90

Reviews of Cohort Studies

Max 85

Cohort Studies

Max 72

Reviews of Case-Control Studies

Max 63

Case-Control Studies

Max 58

Cross-Sectional & Case Series

Max 50

Expert Opinion

Max 5
StrongerWeaker
Randomized Trials
Level 1b
84

84 / 100

Probability of being correct

Participants are randomly assigned to treatment or control groups, minimizing bias. The gold standard for testing whether an intervention causes an effect.

This design can establish causation. Although this is a randomized controlled trial, the lack of blinding and low adherence to the personalized intervention may reduce confidence in causal inference. The study was also not powered for glycemic outcomes, and baseline measurements occurred before randomization, which introduces potential selection bias.

Major COI

Major conflicts that significantly reduce study credibility

The study used a precision nutrition algorithm developed by a co-author (Eran Segal) affiliated with the Weizmann Institute of Science, which was central to the intervention; while no direct industry funding is stated, the algorithm's commercial potential and its exclusive use in the study raise significant conflict concerns.

Conflict Details

Eran Segal
Patent
Employee

Weizmann Institute of Science: Co-inventor of the personalized machine learning algorithm used in the study to predict postprandial glucose response.

The algorithm underpinning the personalized diet arm was developed by a study author and is central to the intervention, creating a clear conflict of interest. Although no industry funder is named, the algorithm has commercial implications and was processed in Israel, raising concerns about data control and potential bias. No COI statement or funding disclosure is present in the provided text.

Key takeaways

  1. 01

    Both groups lowered their blood sugar swings by about 0.8 mg/dL per month and HbA1c by 0.02% per month — no difference between the app group and the simple diet group.

  2. 02

    These changes are statistically real but too small to matter much for daily health — like losing a few grams of sugar over six months.

Surprising findings

  • The personalized group logged food more often than the standard group (43.3% vs. 33.1% of days), yet still saw no better outcomes.You’d expect that more engagement with the app would lead to better results — but even when users followed the app more, it didn’t improve blood sugar control.
  • Both groups reduced HbA1c by only 0.01–0.02% per month — a change too small to be clinically meaningful.Many assume even small HbA1c drops are significant, but this study shows that in a 6-month trial, these changes are statistically real but practically negligible — like losing a few grams of sugar over half a year.

Practical takeaways

If you have prediabetes or type 2 diabetes, focus on consistent calorie control and behavioral support — not expensive apps.

This study only tested one specific algorithm and one type of low-fat diet; results may vary with different interventions or populations.

high confidence

If you’re using a glucose-tracking app, ask yourself: Am I logging food often enough to get feedback? If not, it’s probably wasting your money.

Low adherence (33–43% of days) was common — so even if the app works, it only works if you use it consistently.

medium confidence

Why this study matters

The App Didn’t Outperform a Simple Diet

Both groups — one using a machine learning app that predicted glucose spikes and one on a standard low-fat diet — saw nearly identical drops in glycemic variability (0.83 vs. 0.79 mg/dL per month) and HbA1c (0.02% vs. 0.01% per month), with no statistically significant difference (P=0.92 and P=0.83).

People are spending hundreds on personalized nutrition apps, but this rigorous trial shows they offer no extra benefit over basic calorie restriction and simple dietary guidance.

Low Adherence Killed the Algorithm’s Potential

Participants only logged ≥50% of their calories on 33–43% of days — meaning the app’s personalized feedback was rarely used. The personalized group logged slightly more (43.3%) than the standard group (33.1%), yet still saw no benefit.

No matter how smart the algorithm, if users don’t consistently track food, the tech is useless — highlighting that behavior change matters more than tech precision.

The Real Hero? Calorie Restriction, Not Personalization

Both groups followed identical calorie-restricted plans (500 kcal/day deficit) and received the same behavioral counseling. The only difference was the app’s feedback — yet both improved similarly, suggesting energy balance is the dominant driver.

This flips the narrative: it’s not ‘what you eat’ that matters most for blood sugar — it’s ‘how much you eat.’

The Algorithm Was Built for Israelis — Not Americans

The machine learning model was trained on Israeli microbiome and glucose data, then applied to a U.S. cohort — but the study notes this limits generalizability due to differences in diet, gut bacteria, and demographics.

A tech product marketed as ‘personalized for you’ was trained on a population that doesn’t look like most of its users — raising red flags about AI bias in health tech.

Want the whole report?

Detailed mode opens the full scientific breakdown — every score component, the methodology, conflicts of interest, the evidence analysis behind each claim, and the raw study data.