Machine Learning Predicts Gleevec Nonadherence in Patients With GIST

Although nonadherence to treatment with Gleevec (imatinib) is common in patients with gastrointestinal stromal tumors (GIST) and is associated with poor prognosis and financial burden, investigators found that 53.4% of pooled Chinese patients demonstrated nonadherence, according to research published in Cancer. The research also highlighted that predominant indicators for Gleevec nonadherence were cognitive functioning, whether to perform therapeutic drug monitoring (if_TDM), global health status score, social support and gender.

Based on this data, investigators developed a model based on machine learning (ML) and deep learning (DL) and the goal was to identify factors which predict the risk of Gleevec nonadherence in patients with GIST. The study cites that this is because ML models offer valuable insights which can be gained into factors influencing adherence.

“This study represents the first real‐world investigation using ML techniques to predict risk factors associated with [Gleevec] nonadherence in patients with GIST. By highlighting the potential factors and identifying high‐risk patients, the multidisciplinary medical team can devise targeted strategies to effectively address the daily challenges of treatment adherence,” lead study author, Dr. Li Liu of Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, in Wuhan, China, and co-authors wrote.

Understanding the Methods Utilized in the Study

GIST, which is often driven by KIT or PDGFRA mutations, is effectively managed with Gleevec, a tyrosine kinase inhibitor approved for advanced, metastatic and relapsed/refractory disease. While Gleevec has aided in turning GIST into a manageable condition, nonadherence remains a critical issue due to factors such as side effects, cost and false perceptions of well-being following rapid clinical response or surgery. According to study authors, previous studies have shown high rates of nonadherence, with predictors including gender, income, rural living, adverse reactions, and long-term treatment duration; ML and DL can assist in analyzing complex factors affecting adherence, enabling early identification of high-risk patients and guiding interventions to improve outcomes and quality of life.

To address these unmet needs in care, investigators launched a study at Union Hospital, Tongji Medical College, Huazhong University of Science and Technology. From March 2018 to August 2023, patients who visited the specialized GIST clinic were included and selectively recruited. Eligible patients who were recruited included diagnosis of GIST, being aged 18 years or older and having received Gleevec therapy for at least one month. However, patients were excluded if they had more than 20% incomplete information.

To assess patients on the study, participants partook in a composite questionnaire with the assistance of trained pharmacists. Each interview lasted approximately 45 minutes and consisted of four sections including basic characteristics, adherence assessment, quality of life and social support. Demographic information including age, gender, place of residence, education status, marital status, paid work, monthly income and whether to perform therapeutic drug monitoring was covered. Treatment‐related details like duration of Gleevec therapy, medication reminder, and the number of concomitant medications were available per patient records.

Investigators assessed adherence by the Chinese version of the eight‐item Morisky Medication Adherence Scale (MMAS‐8), which comprises seven items answered with “yes” or “no” alternatives, along with one item rated on a 5‐point Likert scale, with the total score ranging from 0 to 8 and divided into two levels: adherence (score 8), and nonadherence (score less than 8). Furthermore, investigators evaluated quality of life using the Chinese version of the European Organization for Research and Treatment of Cancer Quality of Life Questionnaire C30, which is a 30‐item questionnaire including five functional dimensions, nine symptom dimensions and one measuring Global Health Status. Each raw score is linearly transformed into a standard score ranging from 0 to 100 where higher scores on functional scales represent greater functioning, whereas higher scores on symptom scales represent severe symptoms corresponding to the scale.

Lastly, investigators evaluated social support and assessed this through the validated Chinese Social Support Rating Scale, which consisted of 10 items: three items for objective support, four items for subjective support and the remaining for support utilization; the total score ranges from 12 to 66. Higher scores represent higher levels of social support.

“A total of 397 patients were recruited in this study… Among them, palliative treatment with [Gleevec] was administered to 70 patients [17.6%], whereas the remaining patients received adjuvant therapy. The median duration of [Gleevec] therapy in this population was 19 months. Additionally, the majority of patients had concomitant medications [52.9%],” study authors wrote about the patients who were enrolled.

Delving Further into Data and Key Takeaways

In total, 212 patients (53.4%) were identified as nonadherent to Gleevec therapy. Notably, feeling burdened by the treatment regimen was the most frequently reported reason for nonadherence, comprising 30.73% of the responses.

When establishing the ML model, the screening process incorporated 64 covariates, including basic patient characteristics, quality of life and social support. Nineteen variables were identified as statistically significant and subsequently used for modeling. In total, the study utilized six ML and four DL algorithms for model development. Notably, the light gradient boosting machine (LGBM) model outperformed others. As a result, the LGBM model was selected as the predictive tool for assessing Gleevec adherence in patients with GIST.

However, the study authors noted that, “among 46 patients with adherence, the prediction misclassified 11 as nonadherent, yielding a precision rate of 81.4% and a recall rate of 76.1%. Among the group of 34 nonadherent patients, the prediction incorrectly classified eight as adherent, resulting in a precision rate of 70.2% and a recall rate of 76.5%.”

The scores of variables were calculated and ranked using the LGBM model, with higher scores indicating a stronger impact on Gleevec adherence. Cognitive functioning showed significantly greater importance than other variables, followed by if_TDM and global health status. To further explore the relative importance of factors influencing Gleevec adherence, the Shapley Additive Explanations (SHAP) approach was employed to interpret the positive and negative correlations of variables within the prediction model and the magnitude of their effects. The analysis revealed positive correlations of if_TDM and gender with Gleevec adherence, while pain and insomnia showed negative associations.

A subgroup analysis was performed based on the top five categorical variables: if_TDM, gender, pain, working status and place of residence. Investigators found that patients undergoing Gleevec TDM were more likely to adhere to treatment than those not receiving TDM; male patients demonstrated a higher probability of adherence compared to female patients; patients without pain were more adherent than those with pain levels of 1 or 2, employed or retired individuals were more likely to adhere than unemployed patients and urban residents had a higher likelihood of adherence compared to rural residents.

“An important finding of this study was the significant association between if_TDM and adherence. The results are similar to our previous finding that poor nonadherence was associated with a relatively lower [Gleevec] level. In the present study, [Gleevec] TDM in patients can serve as an objective method for assessing adherence to medication. By continuously monitoring [Gleevec] concentrations, physicians can ascertain whether patients are taking their medication as directed, thereby overcoming limitations of self‐reported adherence measures,” study authors emphasized, concluding that, “Integration of [Gleevec] TDM data with ML and DL techniques has the potential to enhance predictive models by adding a quantitative measure of adherence. This can improve the accuracy of adherence predictions and enable more precise identification of factors contributing to nonadherence.”

For more news on cancer updates, research and education, don’t forget to subscribe to CURE®’s newsletters here.

Reference:

“Imatinib adherence prediction using machine learning approach in patients with gastrointestinal stromal tumor” by Dr. Li Liu, et al., Cancer.