Predicting adherence of imatinib in gastrointestinal stromal tumor—rise of the machines!

Imatinib has transformed the treatment paradigm of gastrointestinal stromal tumors (GISTs), a rare form of mesenchymal tumour that arise in gastrointestinal tract due to activating mutations in receptor tyrosine kinases such as KIT or platelet-derived growth factor receptor alpha (PDGFRA) (1). These sarcomatoid tumours are inherently resistant to traditional chemotherapy unlike other malignant sarcomas. Imatinib has improved the quality of life, progression free survival and overall survival significantly in patients with metastatic GIST, with median overall survival ranging from 4 to 5 years and 10-year overall survival rates reaching around 20% (2). Subsequent trials have also established the benefit of adjuvant imatinib after curative resection. The recommended dose of imatinib in GIST is 400–600 mg once daily without interruption.

However, non-adherence to imatinib is common and well documented in GIST and chronic myeloid leukaemia (CML). We first learnt this from the ADAGIO study which reported one third of the patients were considered non adherent after structured interview (3). Only 14% were considered to be 100% adherent to imatinib. Investigators importantly also showed that patients with suboptimal responses had higher rates of non-adherence. This fact was reproduced in a number of publications in patients with CML as well as GIST (4-6). In the SSGXVIII adjuvant study, 26% (51/198) of patients in the 3-year arm discontinued therapy early for reasons other than recurrence. The most common reason was adverse events (27/51), followed by patient preference (11/51) (7). Not surprisingly, non-adherence is associated with not only poor health outcome but also higher health care cost (8). Then one would have to wonder what leads to such high rates of poor compliance in otherwise a life changing therapy! Medication compliance or noncompliance is not a new concept and not just restricted to imatinib or any anti-cancer therapy. Non-compliance issues are encountered frequently in day-to-day practice and is more common with long term use and polypharmacy but literacy, socio-economic status, family support, drug toxicity and financial toxicity, are some of the well-known factors that affect compliance (3).

Compliance is not always easy to identify especially with increasingly complex use of medications. Moreover, improvements in cancer survivals, have resulted in more patients being followed up in clinic over longer periods which limits the time that can be spent with each patient in clinic. Therefore, it is indeed useful to have tools that could expedite the identification of issues such as non-adherence without compromising clinic time. This study by Liu et al. has tried to identify patients who are at high risk of non-adherence by the use of machine learning (ML) and deep learning (DL) in a single centre Chinese cohort of GIST patients (9). Artificial intelligence (AI) has revolutionised multiple sectors such as technology, transport etc. AI has been used in the health sector, starting from disease screening to clinical decision making, scribing and indeed assisting with precision surgeries.

Although the approach is not novel, the investigators have applied AI in a thoughtful manner to predict non-adherence among patients receiving long-term therapy. This strategy has the potential to reduce clinician workload and, more importantly, to improve patient outcomes.

Koesmahargyo et al. showed that machine learning accurately predicted drug compliance in subsequent day and subsequent week in their clinical trial (10). Others have also attempted to use machine learning to predict adherence in patients with chronic conditions such as acquired immunodeficiency syndrome (AIDS) and heart failure (11,12). In this study by Liu et al. non-adherence was assessed by Morisky Medication Adherence Scale (MMAS) score and somewhat surprisingly non-adherence was quite high in this cohort of patients. Ten AI models were used with an aim of predicting the non-adherence with maximum accuracy. AI model’s prediction performance was measured by precision, accuracy, recall, sensitivity, specificity, F1 score and area under the curve (AUC). They used 80% of the patients to train the AI and then validated these in the other 20% cohort. Light gradient boosting machine (LGBM) algorithm prevailed as the most accurate model with an F1 score and accuracy rate of 0.79 and 0.76, respectively. Sixty-four variables were involved in the screening process and 19 of them were found to be significant after the univariate analysis. Machine learning calculated the “importance score” and author identified 10 covariates as the most important factor contributing to non-adherence. Cognitive impairment topped the list, which is not a surprise, but what is intriguing is that median age of the population in the study was 58 years with a range of 50–68 years.

As anticipated, cognitive impairment is likely to be less prevalent in this age group; however, potential confounding effects of factors such as literacy or educational level should be considered. Furthermore, while the questionnaire was administered by a pharmacist over 45 minutes, the specific method used to assess cognitive function is not clearly described. Given that cognition emerged as the most influential variable in the model, future iterations would benefit from the inclusion of a validated, standardised cognitive assessment tool.

Therapeutic drug monitoring (TDM) of imatinib was one of the ten in the list of importance score. TDM is an interesting discussion point in view of optimizing responses, however it is not widely adopted in clinical practise. A number of studies have looked at and compared the efficacy of 800 vs. 400 mg of imatinib in GIST but most of these studies found no difference in response rates and survival (13). The B2222 study reported, in those with trough plasma concentrations in the lowest quartile (Q1 ≤1,100 µg/L) tended to have lower rates of objective response and a significantly shorter time to progression (TTP =11.3 vs. >30 months, P=0.0029) than patients in the higher quartile groups (Q2 to Q4 >1100 µg/L) (14). TDM may also be more useful in patients who carry a KIT exon 9 mutation where a higher plasma level of Imatinib may help to achieve optimal response (13). Furthermore, there are reports that declining plasma levels of imatinib over time due to reasons that are not well established, but some authors claim this may be due to increased imatinib clearance or poor compliance (15). TDM may be applicable in this setting to maintain adequate dosing.

As identified in this paper, TDM can be a useful tool to optimise the dosing in those patients who are non-compliant, those who experience toxicity and of course, in non-responders. Sixty-one percent of patients in the adherence group underwent TDM whereas only 25% had TDM in the non-adherence group, however it is not clear whether TDM resulted in changes in the dose or frequency of imatinib administration. Although TDM appears to be useful tool to optimise dose in order to increase response rate or mitigate side effects, there are number of constraints such as cost, availability and exact pharmacokinetic parameter (total vs. free imatinib level) which will limit widespread application of its use.

Monthly income was a statistically significant covariate in the study and achieved high importance on AI prediction, ranked 10^th. This is despite the fact that most patients in this cohort received the medication through their medical insurance or rural cooperative medical system whereas only less than 5% of the patients required to self-fund imatinib. This may become a more important limiting factor in jurisdictions where more patients are required to self-fund.

Finally, a number of other covariates that were significant in the univariate analysis that did not reach the top 10 on the importance score; fatigue, insomnia, appetite loss, dyspnoea, physical functioning. These factors are likely to represent side effects of imatinib or symptoms of cancer. In clinical practice, toxicity is an important contributor to non-adherence, even though this was not categorised in the top ten.

The identified factors are well-recognised determinants of non-adherence. Machine-learning approaches enable the efficient integration of multiple variables to accurately identify patients at high risk, which is especially valuable in oncology settings with growing workload pressures and limited resources.

To obtain more generalization data it would be necessary to widen the range of population to include other countries, healthcare systems and other ethnic groups. Furthermore, newer variables can be added on to existing AI models with ease, unlike having to reproduce complex statistical models to incorporate or validate one variable.

The investigators are to be commended for the novel approach and incorporation of AI; however, the practicality of implementing targeted solutions for all the factors identified by the model remains uncertain. Nevertheless, such an efficient AI-based approach to problem identification and data collection is likely to be useful, as optimisation of even a subset of modifiable factors may lead to improved patient outcomes.

Acknowledgments

None.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Journal of Gastrointestinal Oncology. The article did not undergo external peer review.

Funding: None.

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-1-0145/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Demetri GD, von Mehren M, Antonescu CR, et al. NCCN Task Force report: update on the management of patients with gastrointestinal stromal tumors. J Natl Compr Canc Netw 2010;8:S1-41; quiz S42-4. [Crossref] [PubMed]
2. Casali PG, Zalcberg J, Le Cesne A, et al. Ten-Year Progression-Free and Overall Survival in Patients With Unresectable or Metastatic GI Stromal Tumors: Long-Term Analysis of the European Organisation for Research and Treatment of Cancer, Italian Sarcoma Group, and Australasian Gastrointestinal Trials Group Intergroup Phase III Randomized Trial on Imatinib at Two Dose Levels. J Clin Oncol 2017;35:1713-20. [Crossref] [PubMed]
3. Noens L, van Lierde MA, De Bock R, et al. Prevalence, determinants, and outcomes of nonadherence to imatinib therapy in patients with chronic myeloid leukemia: the ADAGIO study. Blood 2009;113:5401-11. [Crossref] [PubMed]
4. Marin D, Bazeos A, Mahon FX, et al. Adherence is the critical factor for achieving molecular responses in patients with chronic myeloid leukemia who achieve complete cytogenetic responses on imatinib. J Clin Oncol 2010;28:2381-8. [Crossref] [PubMed]
5. Pandey D, Chugh P, Chaudhry S. Identifying Key Factors Influencing Non-adherence to Imatinib Therapy in Patients With Chronic Myeloid Leukemia. Cureus 2025;17:e79831. [Crossref] [PubMed]
6. Mazzeo F, Duck L, Joosens E, et al. Nonadherence to imatinib treatment in patients with gastrointestinal stromal tumors: the ADAGIO study. Anticancer Res 2011;31:1407-9. [PubMed]
7. Joensuu H, Eriksson M, Sundby Hall K, et al. One vs three years of adjuvant imatinib for operable gastrointestinal stromal tumor: a randomized trial. JAMA 2012;307:1265-72. [Crossref] [PubMed]
8. Tetzlaff ED, Davey MP. Optimizing adherence to adjuvant imatinib in gastrointestinal stromal tumor. J Adv Pract Oncol 2013;4:238-50. [PubMed]
9. Liu L, Yu Z, Chen H, et al. Imatinib adherence prediction using machine learning approach in patients with gastrointestinal stromal tumor. Cancer 2025;131:e35548. [Crossref] [PubMed]
10. Koesmahargyo V, Abbas A, Zhang L, et al. Accuracy of machine learning-based prediction of medication adherence in clinical research. Psychiatry Res 2020;294:113558. [Crossref] [PubMed]
11. Yeneakal KA, Teferi GH, Mihret TT, et al. Predicting antiretroviral therapy adherence status of adult HIV-positive patients using machine-learning Northwest, Ethiopia, 2025. BMC Med Inform Decis Mak 2025;25:259. [Crossref] [PubMed]
12. Karanasiou GS, Tripoliti EE, Papadopoulos TG, et al. Predicting adherence of patients with HF through machine learning techniques. Healthc Technol Lett 2016;3:165-70. [Crossref] [PubMed]
13. Yip D, Zalcberg J, Ackland S, et al. Controversies in the management of gastrointestinal stromal tumors. Asia Pac J Clin Oncol 2014;10:216-27. [Crossref] [PubMed]
14. Demetri GD, Wang Y, Wehrle E, et al. Imatinib plasma levels are correlated with clinical benefit in patients with unresectable/metastatic gastrointestinal stromal tumors. J Clin Oncol 2009;27:3141-7. [Crossref] [PubMed]
15. Eechoute K, Fransson MN, Reyners AK, et al. A long-term prospective population pharmacokinetic study on imatinib plasma concentrations in GIST patients. Clin Cancer Res 2012;18:5780-7. [Crossref] [PubMed]