Personalized Glycemic Predictions: What the Studies Show

Step 1: show that responses differ

Step 2: build and test a predictor

• Israeli cohort. The machine-learning model predicted glucose responses to real meals and was checked in an independent 100-person cohort. [1]
• Midwestern US cohort (327 people). The Israeli-trained model predicted responses better (R = 0.596) than carbohydrate content (R = 0.395) or calories (R = 0.336), and did slightly better when trained on both cohorts (R = 0.618). [4] An erratum exists. [5]
• UK and US check. A model predicted glucose responses with r = 0.77 and triglyceride responses with r = 0.47, validated in a US cohort of 100. [2]

Prediction accuracy is not the same as health benefit. An R of about 0.6 to 0.8 still leaves much of a person's response unexplained.

Step 3: test a diet built on the predictor

• Prediabetes (225 adults, 6-month intervention). A prediction-guided diet reduced time above 140 mg/dL by 1.3 hours/day versus 0.3 on a Mediterranean diet, and HbA1c fell 0.16% versus 0.08%. [6] The two groups did not differ in a CGM-measured oral glucose tolerance test. [6]
• Newly diagnosed type 2 diabetes (23 adults, pilot). Crossover diets lasted two weeks each, with a 6-month extension in 16 people; the abstract reports lower glucose measures and, at six months, diabetes "remission" by HbA1c below 6.5% in 61%. [7] Small, uncontrolled extension; read cautiously.
• US program (347 adults, 18 weeks). Triglycerides fell more than with standard advice (-0.13 mmol/L), but LDL cholesterol, glucose and insulin did not differ between groups. [8]
• Bread trial. In 2017, average effects of two breads did not differ, but individual responses did. [9]
• Summary review. A 2024 review of seven trials (873 people) in prediabetes or metabolic syndrome found consistent gains in HbA1c and waist circumference and inconsistent results for lipids, blood pressure and fasting glucose. [10]

Limits

• Several trials were short or small, and a lower glucose reading is a biomarker, not a clinical outcome such as diabetes onset or heart disease. See Surrogate vs Patient-Important Outcomes.
• Studies came from a limited number of groups and populations, and the comparison was usually another structured diet, so effects of personalization are tangled with effects of coaching and monitoring.
• The program in one trial mixed several inputs, so the contribution of each is unknown. [8]
• Longer outcomes: see Long-Term Outcomes in Personalized Nutrition Trials.