Study analysis · Frontiers in Artificial Intelligence · 2026
Your starting weight might predict if your GLP-1 drug will work — and it’s not what you think.
Computers can guess who will lose weight or control blood sugar on GLP-1 drugs just by looking at their starting weight and blood sugar levels.
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 looked at a bunch of people who took a medicine and found that certain things—like how much they weighed or how long they had diabetes—were connected to whether the medicine worked well for them. But it didn’t change anything or make people take the medicine differently, so we can’t say those things caused the results.
What’s the bottom line?
Scientists used computer programs to look at people's health data before they started GLP-1 drugs and tried to guess who would lose weight or get their blood sugar under control.
How strong is this study?
The researchers used a big, diverse group of people and tried lots of computer tricks to find patterns, which is pretty smart. But since they just looked back at old records instead of testing things on purpose, we can’t be totally sure their findings will work for everyone else—kind of like guessing the weather from last year’s data.
40 / 100
- COI disclosure+40/40
- Data availabilitydata not shared
- Code availabilitycode not shared
38 / 100
- Randomizationnot randomized
- Blindingblinding unclear
- Control groupno control group
- Sample size (n=11420)+20/20
- Follow-up+10/10
100 / 100
54 / 100
- P-valuesno p-values reported
- Effect size+20/20
- Confidence intervals+15/15
- Pre-registrationnot pre-registered
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 50 / 100
Probability of being correct
Based on clinical experience or non-systematic literature reviews. The lowest level of evidence as they are most susceptible to bias and personal perspective.
This design cannot establish causation — the findings describe an association, not a cause. This is a retrospective observational study with no randomization or control group. It identifies associations between patient characteristics and outcomes but cannot rule out confounding factors or establish that one variable causes another.
No Conflicts
No conflicts of interest identified
No conflicts of interest or funding disclosures were reported in the study.
The study uses data from the All of Us Research Program, a publicly funded initiative, and no industry funding, author affiliations with pharmaceutical companies, or funder involvement in study design, analysis, or publication are disclosed. The absence of a COI or funding statement does not imply conflict, but transparency is limited.
Key takeaways
- 01
The computer was 94% accurate at guessing who would lose enough weight to get below BMI 30, and 79% accurate at guessing who would get HbA1c below 7%.
- 02
People with lower starting weight and BMI were more likely to lose weight.
- 03
People with shorter diabetes duration and lower HbA1c were more likely to control blood sugar.
- 04
This means doctors might soon use simple health data to predict who will benefit most from these expensive drugs — helping avoid trial-and-error treatment.
Surprising findings
- Baseline BMI was the #1 predictor of weight loss — even more than age, sex, or medication type.Most people assume the drug itself or lifestyle changes drive results, but the study shows your starting point — not your effort — is the strongest signal.
- Excluding baseline BMI from the model still yielded 90% AUC for weight loss prediction.This means the model wasn’t just mathematically 'cheating' by using BMI to predict BMI — it was picking up real biological signals from other data.
- HDL cholesterol was a moderate but consistent predictor of weight loss — not just a side note.HDL is rarely discussed in weight loss contexts — this study shows it’s a hidden metabolic marker tied to GLP-1 success.
Practical takeaways
If you're considering GLP-1 therapy, get your baseline BMI and HbA1c tested — lower numbers mean higher odds of success.
This doesn’t mean you shouldn’t try if your numbers are high — it just means you might need higher doses or longer time to see results.
medium confidenceIf you're on insulin or sulfonylureas, discuss with your doctor whether switching to GLP-1 earlier might improve your chances of glycemic control.
This study is observational — it doesn’t prove causation. Don’t stop or change meds without medical advice.
medium confidenceTrack your HDL levels — higher levels may signal better metabolic health and better response to weight-loss drugs.
HDL is just one piece — diet, exercise, and genetics still matter. Don’t fixate on one number.
low confidenceWhy this study matters
AI Predicts Weight Loss With 94% Accuracy
Machine learning models using baseline BMI and weight predicted who would drop below a BMI of 30 with an AUC of 0.94 — meaning they were 94% accurate at distinguishing responders from non-responders. Random Forest and XGBoost were the top performers.
This means doctors could soon use simple health data to avoid costly trial-and-error with GLP-1 drugs like Ozempic — saving time, money, and frustration for millions.
Higher HDL = Better Weight Loss
Contrary to expectations, higher baseline HDL cholesterol was linked to better weight loss outcomes — suggesting metabolic health, not just fat mass, plays a key role in GLP-1 response.
It’s not just about being overweight — your body’s internal metabolism matters. This flips the script on 'more fat = more to lose' assumptions.
Insulin Users Are Less Likely to Control Blood Sugar
Patients already on insulin or sulfonylureas before starting GLP-1 drugs were significantly less likely to reach HbA1c <7%, indicating these drugs work best in earlier-stage diabetes.
It’s not that GLP-1 drugs don’t work for advanced diabetes — it’s that those patients have more damaged pancreases. This helps explain why some people feel the drugs 'don’t work' for them.
Glycemic Control Is Harder to Predict Than Weight Loss
While weight loss prediction hit AUC 0.94, glycemic control only reached AUC 0.79 — meaning the computer was only 79% accurate at predicting who’d hit HbA1c <7%, revealing more complexity in blood sugar response.
Even with all the data, blood sugar is messier than weight loss — hinting that lifestyle, adherence, or unknown factors play a bigger role than we thought.
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 used computer programs to look at people's health data before they started GLP-1 drugs and tried to guess who would lose weight or get their blood sugar under control.
Research results
The computer was 94% accurate at guessing who would lose enough weight to get below BMI 30, and 79% accurate at guessing who would get HbA1c below 7%. People with lower starting weight and BMI were more likely to lose weight. People with shorter diabetes duration and lower HbA1c were more likely to control blood sugar.
What this means - more context
This means doctors might soon use simple health data to predict who will benefit most from these expensive drugs — helping avoid trial-and-error treatment.
This study aimed to develop machine learning models to predict glycemic control and weight loss outcomes in adults with type 2 diabetes starting GLP-1 receptor agonist therapy, using real-world data to identify key predictors of response.
Tree-based ensemble models (RF, XGBoost) achieved high accuracy in predicting weight loss (AUC ≈ 0.94) and moderate accuracy in predicting glycemic control (AUC ≈ 0.79) using baseline clinical data. Baseline BMI and weight were strongest predictors of weight loss; duration of diabetes, baseline HbA1c, and sulfonylurea/insulin use were key for glycemic control. Higher baseline HDL was associated with better weight loss outcomes.
Methods Used
Retrospective cohort study using data from the All of Us Research Program (n=11,420 for weight loss, n=3,975 for glycemic control). Machine learning models (logistic regression, random forest, XGBoost, SVM, neural networks, LightGBM, CatBoost) were trained and validated using 10-fold cross-validation. SHAP analysis was used for feature importance interpretation.
Main Finding
Random forest and XGBoost models achieved AUC of ≈0.94 for predicting weight loss (BMI <30 kg/m²) and ≈0.79 for predicting glycemic control (HbA1c <7%). Baseline BMI and weight were the top predictors for weight loss; duration of diabetes, baseline HbA1c, and sulfonylurea/insulin use were top predictors for glycemic control.
Confidence Level
High internal validity due to large sample size, rigorous cross-validation, and use of explainable AI (SHAP); confidence limited by retrospective design, lack of external validation, and absence of adherence or dose data.
Study Flags
Red Flags
- •Retrospective design limits causal inference
- •No external validation performed
- •Lack of data on medication adherence or dose escalation
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
Baseline BMI was the #1 predictor of weight loss — even more than age, sex, or medication type.
Most people assume the drug itself or lifestyle changes drive results, but the study shows your starting point — not your effort — is the strongest signal.
Practical Takeaways
If you're considering GLP-1 therapy, get your baseline BMI and HbA1c tested — lower numbers mean higher odds of success.
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 50 / 100
Probability of being correct
Based on clinical experience or non-systematic literature reviews. The lowest level of evidence as they are most susceptible to bias and personal perspective.
Non-Scorable
Subject
Lower probability
on the GRADE evidence scale
This study looked at a bunch of people who took a medicine and found that certain things—like how much they weighed or how long they had diabetes—were connected to whether the medicine worked well for them. But it didn’t change anything or make people take the medicine differently, so we can’t say those things caused the results.
No conflicts of interest were detected in this study. No score impact.
Strengths
- Large sample size (n=11,420) with diverse population from the All of Us Research Program
- Use of multiple machine learning algorithms with rigorous 10-fold cross-validation
- Application of SHAP for interpretability and feature importance analysis
Weaknesses
- Retrospective design with no randomization or control group
- Blinding status unknown, increasing risk of selection and measurement bias
- Potential for unmeasured confounding (e.g., diet, exercise, adherence)
Methodology
Evidence Keywords
Statistical Reporting
Not medical advice. For informational purposes only. Always consult a healthcare professional. Terms
Scientists used computer programs to look at people's health data before they started GLP-1 drugs and tried to guess who would lose weight or get their blood sugar under control.
Research results
The computer was 94% accurate at guessing who would lose enough weight to get below BMI 30, and 79% accurate at guessing who would get HbA1c below 7%. People with lower starting weight and BMI were more likely to lose weight. People with shorter diabetes duration and lower HbA1c were more likely to control blood sugar.
What this means - more context
This means doctors might soon use simple health data to predict who will benefit most from these expensive drugs — helping avoid trial-and-error treatment.
This study aimed to develop machine learning models to predict glycemic control and weight loss outcomes in adults with type 2 diabetes starting GLP-1 receptor agonist therapy, using real-world data to identify key predictors of response.
Tree-based ensemble models (RF, XGBoost) achieved high accuracy in predicting weight loss (AUC ≈ 0.94) and moderate accuracy in predicting glycemic control (AUC ≈ 0.79) using baseline clinical data. Baseline BMI and weight were strongest predictors of weight loss; duration of diabetes, baseline HbA1c, and sulfonylurea/insulin use were key for glycemic control. Higher baseline HDL was associated with better weight loss outcomes.
Methods Used
Retrospective cohort study using data from the All of Us Research Program (n=11,420 for weight loss, n=3,975 for glycemic control). Machine learning models (logistic regression, random forest, XGBoost, SVM, neural networks, LightGBM, CatBoost) were trained and validated using 10-fold cross-validation. SHAP analysis was used for feature importance interpretation.
Main Finding
Random forest and XGBoost models achieved AUC of ≈0.94 for predicting weight loss (BMI <30 kg/m²) and ≈0.79 for predicting glycemic control (HbA1c <7%). Baseline BMI and weight were the top predictors for weight loss; duration of diabetes, baseline HbA1c, and sulfonylurea/insulin use were top predictors for glycemic control.
Confidence Level
High internal validity due to large sample size, rigorous cross-validation, and use of explainable AI (SHAP); confidence limited by retrospective design, lack of external validation, and absence of adherence or dose data.
Study Flags
Red Flags
- •Retrospective design limits causal inference
- •No external validation performed
- •Lack of data on medication adherence or dose escalation
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
Baseline BMI was the #1 predictor of weight loss — even more than age, sex, or medication type.
Most people assume the drug itself or lifestyle changes drive results, but the study shows your starting point — not your effort — is the strongest signal.
Practical Takeaways
If you're considering GLP-1 therapy, get your baseline BMI and HbA1c tested — lower numbers mean higher odds of success.
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 50 / 100
Probability of being correct
Based on clinical experience or non-systematic literature reviews. The lowest level of evidence as they are most susceptible to bias and personal perspective.
Non-Scorable
Subject
Lower probability
on the GRADE evidence scale
This study looked at a bunch of people who took a medicine and found that certain things—like how much they weighed or how long they had diabetes—were connected to whether the medicine worked well for them. But it didn’t change anything or make people take the medicine differently, so we can’t say those things caused the results.
No conflicts of interest were detected in this study. No score impact.
Strengths
- Large sample size (n=11,420) with diverse population from the All of Us Research Program
- Use of multiple machine learning algorithms with rigorous 10-fold cross-validation
- Application of SHAP for interpretability and feature importance analysis
Weaknesses
- Retrospective design with no randomization or control group
- Blinding status unknown, increasing risk of selection and measurement bias
- Potential for unmeasured confounding (e.g., diet, exercise, adherence)
Methodology
Evidence Keywords
Statistical Reporting
Scoring
How strong is this study?
The researchers used a big, diverse group of people and tried lots of computer tricks to find patterns, which is pretty smart. But since they just looked back at old records instead of testing things on purpose, we can’t be totally sure their findings will work for everyone else—kind of like guessing the weather from last year’s data.
40 / 100
- COI disclosure+40/40
- Data availabilitydata not shared
- Code availabilitycode not shared
38 / 100
- Randomizationnot randomized
- Blindingblinding unclear
- Control groupno control group
- Sample size (n=11420)+20/20
- Follow-up+10/10
100 / 100
54 / 100
- P-valuesno p-values reported
- Effect size+20/20
- Confidence intervals+15/15
- Pre-registrationnot pre-registered
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 50 / 100
Probability of being correct
Based on clinical experience or non-systematic literature reviews. The lowest level of evidence as they are most susceptible to bias and personal perspective.
This design cannot establish causation — the findings describe an association, not a cause. This is a retrospective observational study with no randomization or control group. It identifies associations between patient characteristics and outcomes but cannot rule out confounding factors or establish that one variable causes another.
No Conflicts
No conflicts of interest identified
No conflicts of interest or funding disclosures were reported in the study.
The study uses data from the All of Us Research Program, a publicly funded initiative, and no industry funding, author affiliations with pharmaceutical companies, or funder involvement in study design, analysis, or publication are disclosed. The absence of a COI or funding statement does not imply conflict, but transparency is limited.