SMS message sentiment as a predictive indicator of quality of life and program adherence in oncology nutrition.

1627 Background: A text-message based artificial intelligence nutrition platform called “Ina” was previously developed and shown to be equivalent to human dietitians in providing nutritional recommendations to patients with cancer. Once registered on the platform, patient users exchange conversational texts with and receive tailored dietary recommendations and recipes from Ina. We investigated whether the emotional sentiment of SMS messages can act as a proactive predictor of program-related Quality of Life (QoL) and program adherence. Methods: QoL was captured via a 1-5 Likert scale (“Has using Ina improved your quality of life?”) survey, and program adherence was captured by a binary response to a question (“Have you used any of Ina’s suggested tips?”). We extracted SMS messages within a 60-day window prior to each survey and question. Sentiment was measured using the Jockers-Rinker lexicon via the validated sentimentr package; scores were calculated as a weighted average ranging from -1 (negative sentiment) to +1 (positive sentiment). Linear mixed-effects models (LMM) were used to quantify the relationship between pre-survey message sentiment and program QoL/adherence outcomes, adjusting for total number of surveys completed and user’s unique baseline sentiment. Results: Between October 2019 and January 2026, 14,700 unique text messages were analyzed from 540 unique patient users of the platform (mean age: 51± 20 years; 91% female). Cancer types included breast (40%), ovarian (24%), bladder (16%), lung (13%), and colon (7%); metastases were present among 34%. For every one-point improvement in digital sentiment, users experienced a 14% improvement in subsequent program-related QoL (𝛽= 0.72, p < 0.001). Furthermore, there was a trend toward an association between higher positive digital sentiment and subsequent reported adherence to nutritional recommendations (𝛽 = 0.12, p = 0.097). Conclusions: Sentiment analysis of unstructured SMS messages successfully identified subtle shifts in digital tone that precede structured assessments. This 'leading indicator' of sentiment could enable proactive supportive care and provider escalation for timely triage and intervention. Given rapid symptomatic deterioration in oncology that often leads to non-adherence, text-based sentiment analysis could enable early detection of declining engagement to support improved quality of life and long-term outcomes.