When precision medicine meets human behavior: limits of artificial intelligence in predicting imatinib adherence

The introduction of imatinib has transformed gastrointestinal stromal tumor (GIST) from a largely lethal sarcoma into a chronic, manageable disease for many patients (1,2). With prolonged survival, the determinants of long-term treatment success have shifted from molecular sensitivity alone to factors that sustain durable therapy (3,4). One such factor—medication adherence—has emerged as a critical yet underexplored determinant of long-term treatment success (5,6). In this context, the study by Liu et al. (7) represents a timely effort to apply machine learning (ML) and deep learning (DL) approaches to predict imatinib adherence in patients with GIST.

Using a relatively large real-world cohort of 397 patients, the authors constructed and compared multiple ML and DL models. They identified cognitive impairment, absence of therapeutic drug monitoring (TDM), poorer overall health status, lower scores on the Social Support Rating Scale (SSRS), and female sex as factors associated with an increased risk of imatinib non-adherence. This study is among the first to apply ML- and DL-based approaches to medication adherence in GIST, by attempting to address determinants of treatment continuity that are not captured by molecular profiling alone. However, despite its conceptual appeal, several methodological and interpretative limitations substantially constrain the conclusions that can be drawn.

A central concern lies in the definition of adherence itself. The authors dichotomized adherence using the Morisky Medication Adherence Scale-8 (MMAS-8), classifying only patients with a perfect score of 8 as adherent and all others as nonadherent. Although this cutoff has precedent in the literature (8), it represents an exceptionally stringent definition. Under this scheme, even a single missed dose or minor deviation categorizes a patient as nonadherent. While such rigor may be defensible for exploratory modeling, its clinical meaning remains unclear. Crucially, the study does not demonstrate whether this binary classification correlates with clinically meaningful outcomes, such as recurrence-free survival, treatment discontinuation, or the emergence of resistance.

In GIST, where adjuvant imatinib is typically prescribed for 3 years and therapeutic benefit depends on sustained exposure (4,9), the clinical implications of occasional nonadherence are likely to differ substantially from those of persistent nonadherence, including premature discontinuation of therapy. By collapsing all deviations into a single “nonadherent” category, the model may obscure these distinctions. Without linkage to oncologic outcomes, it remains uncertain whether the predicted “nonadherence” identified by the model represents a clinically actionable risk or merely behavioral noise. As a result, the model predicts adherence to a questionnaire-based construct rather than to a therapeutic endpoint that matters to patients and clinicians.

An additional concern pertains to the validation of the adherence instrument. Liu et al. cite the 2008 validation study of the MMAS-8 to justify its use. However, that validation article was formally retracted in 2023 by The Journal of Clinical Hypertension (10). Although the MMAS-8 continues to be widely used, the retraction of its foundational validation study inevitably raises questions regarding the robustness of the adherence construct serving as the supervisory label in predictive modeling. This issue warrants explicit acknowledgment.

A second limitation concerns the use—and perhaps underuse—of artificial intelligence itself. Despite deploying multiple ML and DL algorithms, the analytic framework ultimately relies on conventional variable selection through univariate screening, followed by modeling on a modest number of predefined covariates. The strongest predictors identified—cognitive functioning, social support, global health status, and sex—are well-established, or at least readily inferable in routine clinical practice, determinants of medication adherence across chronic diseases. In this sense, the findings are intuitive rather than revelatory. More importantly, the promise of DL lies in its ability to discover latent, nonlinear patterns from high-dimensional or unstructured data, often without explicit feature engineering (11). In the present study, however, DL did not yield fundamentally new insights beyond those obtainable with traditional statistical approaches. Indeed, the best-performing model, light gradient boosting machine (LGBM), is a tree-based ML method rather than a deep neural network. The DL models included did not clearly outperform simpler approaches, raising the question of whether artificial intelligence meaningfully advanced understanding in this setting or merely provided a different computational lens for established associations.

This limitation is reflected in the model’s predictive performance. Although the reported area under the receiver operating characteristic curve of 0.81 is respectable, overall accuracy and F1-scores remained in the moderate range. Such performance is unlikely to support individual-level clinical decision-making and reinforces the authors’ implicit positioning of the model as a screening or risk stratification tool rather than a definitive predictor. While this is a reasonable goal, it also underscores the gap between methodological sophistication and clinical utility.

Additional issues further temper the interpretation. By virtue of its cross-sectional design, adherence was necessarily assessed at a single time point during ongoing imatinib therapy rather than over the intended 3-year adjuvant course. Consequently, patients who discontinued treatment early because of severe toxicity or profound nonadherence, as well as those who discontinued therapy after the time of assessment, may not have been adequately represented, introducing potential survivorship or selection bias. Moreover, adherence was assessed exclusively through self-reported questionnaires, which are vulnerable to recall and social desirability biases and may further limit model performance.

Despite these caveats, the study does offer several important conceptual contributions. The strong influence of cognitive functioning highlights the cognitive demands of long-term oral anticancer therapy—an aspect often underestimated in oncology practice. The association between TDM and adherence is particularly noteworthy, suggesting that objective pharmacokinetic monitoring may serve not only as a dosing tool but also as a behavioral surrogate for adherence. This finding points toward a promising future direction in which quantitative drug exposure data are integrated with predictive models to improve both accuracy and interpretability.

More broadly, the study reflects an evolving paradigm in oncology: as targeted therapies render cancers biologically controllable, the limiting factor increasingly becomes the patient’s capacity to sustain treatment. From this perspective, the modest performance of the model may be less a failure of artificial intelligence than a reminder of the inherent complexity of human behavior. Adherence is influenced by psychological, social, and contextual factors that may not be fully captured in structured datasets, even with increasingly sophisticated algorithms.

To meaningfully advance this field, future investigations should consider first establishing a clinically explicit definition of incomplete adjuvant therapy—such as premature discontinuation or failure to complete the intended 3-year course—and then identifying its predictors using large-scale, population-based datasets. Importantly, such analyses should extend beyond conventional biomedical variables to incorporate socioeconomic and contextual determinants, including income level, residential environment, household structure, and type of health insurance coverage. By leveraging high-dimensional data and artificial intelligence to model these multifactorial influences, researchers may more accurately estimate the risk of treatment discontinuation and thereby clarify the true clinical potential of ML in medication adherence research.

In summary, this study represents an important step toward incorporating ML into the evaluation of medication adherence in GIST. Its strengths lie in its real-world focus and its effort to translate predictive modeling into clinically interpretable insights. However, the absence of linkage between adherence definitions and oncologic outcomes, the reliance on predefined variables rather than genuine data-driven discovery, and the modest predictive performance limit its immediate clinical impact. At present, the study should be viewed not as a definitive solution, but as a thoughtful—and appropriately cautious—exploration of what precision care might look like in the era of long-term targeted therapy.

Acknowledgments

None.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Journal of Gastrointestinal Oncology. The article has undergone external peer review.

Peer Review File: Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-1-0062/prf

Funding: None.

Conflicts of Interest: The author has completed the ICMJE uniform disclosure form (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-1-0062/coif). The author has no conflicts of interest to declare.

Ethical Statement: The author is 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. 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]
2. Kanda T, Ichikawa H, Ishikawa T, et al. Fourteen-year follow-up results of imatinib therapy in patients with unresectable and metastatic gastrointestinal stromal tumors. Int J Clin Oncol 2024;29:1870-7. [Crossref] [PubMed]
3. Le Cesne A, Ray-Coquard I, Bui BN, et al. Discontinuation of imatinib in patients with advanced gastrointestinal stromal tumours after 3 years of treatment: an open-label multicentre randomised phase 3 trial. Lancet Oncol 2010;11:942-9. [Crossref] [PubMed]
4. Ushimaru Y, Takahashi T, Nakajima K, et al. Real-world data on the efficacy and safety of adjuvant chemotherapy in Japanese patients with a high-risk of gastrointestinal stromal tumor recurrence. Int J Clin Oncol 2022;27:921-9. [Crossref] [PubMed]
5. von Mehren M, Widmer N. Correlations between imatinib pharmacokinetics, pharmacodynamics, adherence, and clinical response in advanced metastatic gastrointestinal stromal tumor (GIST): an emerging role for drug blood level testing? Cancer Treat Rev 2011;37:291-9. [Crossref] [PubMed]
6. Blay JY, Rutkowski P. Adherence to imatinib therapy in patients with gastrointestinal stromal tumors. Cancer Treat Rev 2014;40:242-7. [Crossref] [PubMed]
7. 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]
8. Wang Y, Zhang P, Han Y, et al. Adherence to Adjuvant Imatinib Therapy in Patients with Gastrointestinal Stromal Tumor in Clinical Practice: A Cross-Sectional Study. Chemotherapy 2019;64:197-204. [Crossref] [PubMed]
9. Joensuu H, Eriksson M, Sundby Hall K, et al. Survival Outcomes Associated With 3 Years vs 1 Year of Adjuvant Imatinib for Patients With High-Risk Gastrointestinal Stromal Tumors: An Analysis of a Randomized Clinical Trial After 10-Year Follow-up. JAMA Oncol 2020;6:1241-6. [Crossref] [PubMed]
10. Morisky DE, Ang A, Krousel-Wood M, et al. Predictive validity of a medication adherence measure in an outpatient setting. J Clin Hypertens (Greenwich) 2008;10:348-54. [Crossref] [PubMed]
11. Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning in healthcare. Nat Med 2019;25:24-9. [Crossref] [PubMed]