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.
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.
75 / 100
- COI disclosure+40/40
- Data availability+35/35
- Code availabilitycode not shared
70 / 100
- Randomization+20/20
- Blindingnot blinded
- Control group+15/15
- Sample size (n=156)+10.8/20
- Follow-up+10/10
100 / 100
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 reviewsReviews of RCTs (Meta-analyses)
Max 100Randomized TrialsRandomized Trials
Max 90Reviews of Cohort StudiesReviews of Cohort Studies
Max 85Cohort StudiesCohort Studies
Max 72Reviews of Case-Control StudiesReviews of Case-Control Studies
Max 63Case-Control StudiesCase-Control Studies
Max 58Cross-Sectional & Case SeriesCross-Sectional & Case Series
Max 50Expert OpinionExpert Opinion
Max 584 / 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
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
- 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.
- 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 confidenceIf 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 confidenceWhy 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.
Overview
What the study found
The study in plain English — the bottom line, every takeaway we extracted, and what to do with them.
Not medical advice. For informational purposes only. Always consult a healthcare professional. Terms
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.
Research results
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.
What this means - more context
These changes are statistically real but too small to matter much for daily health — like losing a few grams of sugar over six months.
This study compared a personalized nutrition diet based on postprandial glucose predictions with a standardized low-fat diet in adults with prediabetes or moderately controlled type 2 diabetes to assess impacts on glycemic variability and HbA1c.
Both groups experienced similar small reductions in glycemic variability (MAGE: ~0.8 mg/dL/month) and HbA1c (~0.02% per month) over six months, with no statistically significant difference between the personalized and standardized diet groups, despite the personalized arm using a machine learning algorithm and continuous glucose monitoring.
Methods Used
Randomized clinical trial with 156 adults (prediabetes or moderately controlled T2D); participants assigned to either a standardized low-fat diet or a personalized diet guided by a machine learning algorithm trained on Israeli microbiome and glucose data; both groups received identical calorie restriction, behavioral counseling, and smartphone-based dietary self-monitoring; glycemic outcomes measured via continuous glucose monitoring (CGM) and HbA1c at baseline, 3, and 6 months using intention-to-treat linear mixed models.
Main Finding
Personalized nutrition did not improve glycemic outcomes compared to standardized calorie-restricted diet; MAGE decreased by 0.83 mg/dL/month (standardized) and 0.79 mg/dL/month (personalized; P=0.92 for between-group difference); HbA1c decreased by 0.02% per month (standardized) and 0.01% per month (personalized; P=0.83).
Confidence Level
Moderate; robust RCT design with intention-to-treat analysis, CGM, and effect sizes reported, but limited by low dietary self-monitoring adherence (33–43%), potential lack of generalizability due to non-representative sample (mostly White, female, high-income), and algorithm trained on Israeli population.
Study Flags
Red Flags
- •Low dietary self-monitoring adherence (33–43% of days)
- •Sample not representative (overrepresented White, female, high-income; underrepresented Hispanic and male)
- •Algorithm trained on Israeli population, limiting generalizability to U.S. populations
No biological mechanisms were identified in this study. This may be an epidemiological, observational, or survey-based study that reports associations rather than proposing causal biological pathways.
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.
Practical Takeaways
If you have prediabetes or type 2 diabetes, focus on consistent calorie control and behavioral support — not expensive apps.
RCT reviewsReviews of RCTs (Meta-analyses)
Max 100Randomized TrialsRandomized Trials
Max 90Reviews of Cohort StudiesReviews of Cohort Studies
Max 85Cohort StudiesCohort Studies
Max 72Reviews of Case-Control StudiesReviews of Case-Control Studies
Max 63Case-Control StudiesCase-Control Studies
Max 58Cross-Sectional & Case SeriesCross-Sectional & Case Series
Max 50Expert OpinionExpert Opinion
Max 584 / 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.
Human RCT
Subject
High probability
on the GRADE evidence scale
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.
Major conflicts — industry funding with significant control over the research. Reporting score and overall score cap have been reduced.
Strengths
- Randomized controlled trial design with intention-to-treat analysis
- Objective outcome measures using continuous glucose monitoring
- Pre-registered trial with published protocol
Weaknesses
- No blinding of participants or researchers
- Baseline measurements occurred before randomization
- Low adherence to dietary self-monitoring (especially in standardized group)
Methodology
Evidence Keywords
Statistical Reporting
Not medical advice. For informational purposes only. Always consult a healthcare professional. Terms
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.
Research results
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.
What this means - more context
These changes are statistically real but too small to matter much for daily health — like losing a few grams of sugar over six months.
This study compared a personalized nutrition diet based on postprandial glucose predictions with a standardized low-fat diet in adults with prediabetes or moderately controlled type 2 diabetes to assess impacts on glycemic variability and HbA1c.
Both groups experienced similar small reductions in glycemic variability (MAGE: ~0.8 mg/dL/month) and HbA1c (~0.02% per month) over six months, with no statistically significant difference between the personalized and standardized diet groups, despite the personalized arm using a machine learning algorithm and continuous glucose monitoring.
Methods Used
Randomized clinical trial with 156 adults (prediabetes or moderately controlled T2D); participants assigned to either a standardized low-fat diet or a personalized diet guided by a machine learning algorithm trained on Israeli microbiome and glucose data; both groups received identical calorie restriction, behavioral counseling, and smartphone-based dietary self-monitoring; glycemic outcomes measured via continuous glucose monitoring (CGM) and HbA1c at baseline, 3, and 6 months using intention-to-treat linear mixed models.
Main Finding
Personalized nutrition did not improve glycemic outcomes compared to standardized calorie-restricted diet; MAGE decreased by 0.83 mg/dL/month (standardized) and 0.79 mg/dL/month (personalized; P=0.92 for between-group difference); HbA1c decreased by 0.02% per month (standardized) and 0.01% per month (personalized; P=0.83).
Confidence Level
Moderate; robust RCT design with intention-to-treat analysis, CGM, and effect sizes reported, but limited by low dietary self-monitoring adherence (33–43%), potential lack of generalizability due to non-representative sample (mostly White, female, high-income), and algorithm trained on Israeli population.
Study Flags
Red Flags
- •Low dietary self-monitoring adherence (33–43% of days)
- •Sample not representative (overrepresented White, female, high-income; underrepresented Hispanic and male)
- •Algorithm trained on Israeli population, limiting generalizability to U.S. populations
No biological mechanisms were identified in this study. This may be an epidemiological, observational, or survey-based study that reports associations rather than proposing causal biological pathways.
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.
Practical Takeaways
If you have prediabetes or type 2 diabetes, focus on consistent calorie control and behavioral support — not expensive apps.
RCT reviewsReviews of RCTs (Meta-analyses)
Max 100Randomized TrialsRandomized Trials
Max 90Reviews of Cohort StudiesReviews of Cohort Studies
Max 85Cohort StudiesCohort Studies
Max 72Reviews of Case-Control StudiesReviews of Case-Control Studies
Max 63Case-Control StudiesCase-Control Studies
Max 58Cross-Sectional & Case SeriesCross-Sectional & Case Series
Max 50Expert OpinionExpert Opinion
Max 584 / 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.
Human RCT
Subject
High probability
on the GRADE evidence scale
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.
Major conflicts — industry funding with significant control over the research. Reporting score and overall score cap have been reduced.
Strengths
- Randomized controlled trial design with intention-to-treat analysis
- Objective outcome measures using continuous glucose monitoring
- Pre-registered trial with published protocol
Weaknesses
- No blinding of participants or researchers
- Baseline measurements occurred before randomization
- Low adherence to dietary self-monitoring (especially in standardized group)
Methodology
Evidence Keywords
Statistical Reporting
Scoring
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.
75 / 100
- COI disclosure+40/40
- Data availability+35/35
- Code availabilitycode not shared
70 / 100
- Randomization+20/20
- Blindingnot blinded
- Control group+15/15
- Sample size (n=156)+10.8/20
- Follow-up+10/10
100 / 100
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 reviewsReviews of RCTs (Meta-analyses)
Max 100Randomized TrialsRandomized Trials
Max 90Reviews of Cohort StudiesReviews of Cohort Studies
Max 85Cohort StudiesCohort Studies
Max 72Reviews of Case-Control StudiesReviews of Case-Control Studies
Max 63Case-Control StudiesCase-Control Studies
Max 58Cross-Sectional & Case SeriesCross-Sectional & Case Series
Max 50Expert OpinionExpert Opinion
Max 584 / 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
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.