Removing dietary triggers can relieve the symptoms of irritable bowel syndrome (IBS) and is as effective as pharmaceutical therapy for managing inflammatory bowel disorders (IBD) in certain populations. However, “trigger foods” are so heterogeneous that nutrition plans must be tailored to each patient’s biological, lifestyle, and clinical characteristics.

That’s a time-consuming, complex endeavor that frustrates patients and can lead to noncompliance with dietary intervention.

Joshua R. Korzenik, MD, director of the Crohn’s and Colitis Center at Brigham and Women’s Hospital, and colleagues have created an interactive, mobile-delivered elimination diet program for IBS and dual IBS/IBD that’s powered by machine learning that delivers precise and personalized recommendations. In Clinical and Translational Gastroenterology, they report encouraging results from a pilot study of the new tool.

Study Design

37 participants, mean age 37 (range, 19–57), were recruited via social media and received financial compensation:

• 16 participants with IBS who rated their symptoms on the IBS Symptom Severity Score (IBS-SSS)

• 12 participants with both IBS and ulcerative colitis who rated their symptoms on the IBS-SSS and the Patient Simple Clinical Colitis Activity Index (P-SCCAI)

• 9 participants with both IBS and Crohn’s disease who rated their symptoms on the IBS-SSS and the Mobile Health Index for CD (mHI-CD)

Based on their systematic literature review, the researchers generated a list of 246 foods that frequently cause gastrointestinal distress. A machine learning algorithm mapped them to clinical and demographic characteristics and derived for each participant a unique set of 21 likely trigger foods.

The Intervention

Every morning for nine weeks, participants received a smartphone text message that linked to a digital survey about symptoms, symptom severity scores, and foods eaten during the previous 24 hours. Machine learning used the responses to guide participants through a four-phase intervention:

• Phase 1—Identification (3 weeks)—Participants ate their regular diet. Machine learning named the three to five trigger foods most strongly associated with each participant’s adverse symptoms

• Phase 2—Elimination (2 weeks)—Participants eliminated those three to five foods from their diet

• Phase 3—Reintroduction (about 1–2 weeks)—Participants were guided to reintroduce eliminated foods one at a time, every three days. Each day they increased their intake of those foods by one serving. If daily symptom scores increased by ≥25%, participants paused reintroduction until symptoms abated to phase 2 baseline values

• Phase 4—Maintenance (2 weeks)—Participants were asked to follow their modified diet (removing the trigger foods identified in earlier phases)

Primary Outcome

The primary outcome was symptomatic improvement, measured in four ways:

• Statistically significant improvement on the relevant symptom scale(s)—81% of participants at week 5 and 70% at week 9

• Clinically significant symptom improvement (≥35-point reduction of IBS-SSS, ≥1-point reduction of P-SCCAI or ≥30% reduction of mHI-CD)—78% of participants at week 5 and 62% at week 9 (P=0.18 for week 5 vs. week 9)

• Total resolution of symptoms (IBS-SSS <75, P-SCCAI <2, and mHI-CD <5.5)—68% of participants overall, 50% of those with IBS, 63% of those with IBS/UC, and 78% of those with IBS/CD; participants were 14 times more likely to have fully resolved symptoms at study end compared with baseline (P<0.001)

• Persistence of clinically significant symptom relief—No significant difference between the study midpoint and endpoint (P=0.15), indicating persistent relief

Secondary Outcomes

The pilot program was judged to be feasible and desirable based on the proportion of daily surveys completed (95%), retention in the study (95% of participants), adherence to the study protocol (89% of participants), percentage reduction of trigger food intake (89%), and satisfaction with the experience (92% of participants).

Differences From Existing Interventions

The digital program offered several key advantages over existing dietary interventions:

• High degree of personalization—No two final diets generated by the algorithm were the same, which emphasizes the need to personalize nutritional interventions to an individual rather than a disease or population

• Less restriction—Personalization allowed participants to eat a more nutritionally diverse set of foods than with elimination diets such as FODMAP

• Adherence—The program’s personalization and digitization contributed to considerably higher levels of compliance (89%) relative to previously reported dietary interventions (16%–50%)