Assessing medication-related burden and medication adherence among older patients from Central Nepal: A machine learning approach

This study utilizes a machine learning approach to analyze medication-related burden and adherence among older patients in Central Nepal, identifying polypharmacy and the need for assistance as key drivers while advocating for targeted clinical interventions like deprescribing.

The Gist

Imagine you are an older adult living in Nepal. You have a few chronic health issues, like diabetes or high blood pressure. To stay healthy, your doctor gives you a "prescription backpack" filled with different pills. Some you take once a day, some twice, some with food, some on an empty stomach.

This study is like a deep dive into how heavy that "prescription backpack" feels to the people carrying it, and whether they are actually able to keep it on their backs every single day.

Here is the story of the research, broken down into simple parts:

1. The Problem: The "Backpack" Gets Too Heavy

As people get older, they often need more medicine to manage their health. But having too many pills creates a Medication-Related Burden (MRB). Think of this burden not just as the weight of the pills, but as the mental stress, the cost, the confusion, and the time it takes to manage them.

The researchers wanted to know:

• How heavy does this backpack feel to older patients in Central Nepal?
• Are they dropping the backpack (stopping meds) or forgetting to put it on (missing doses)?
• What specific things make the backpack feel heaviest?

2. The Method: A New Kind of Detective Work

Usually, doctors use simple math (like adding up numbers) to figure out what causes problems. But human behavior is messy and complicated, like a tangled ball of yarn. Simple math often misses the knots.

So, this team used Machine Learning (AI).

• The Analogy: Imagine you have a huge pile of clues (age, income, number of pills, who helps them, etc.). A human detective might look at one clue at a time. The AI is like a super-smart detective that looks at all the clues at once, sees how they tangle together, and predicts exactly who is struggling the most.

They interviewed 390 older patients at Bharatpur Hospital. They asked them two main questions using special surveys:

1. LMQ-3: "How much of a burden does your medicine routine feel?" (The "Backpack Weight" test).
2. ARMS: "How often do you forget to take your meds or run out?" (The "Did you drop the backpack?" test).

3. The Findings: What the AI Discovered

The results were clear, and the AI was very good at spotting patterns.

The "Backpack" Weight:

• On average, the patients felt a moderate burden. It wasn't crushing, but it was definitely heavy enough to be annoying.
• They trusted their doctors, but they felt stressed about side effects and the practical hassle of getting pills.

The "Dropped Backpack" (Non-Adherence):

• Many patients were moderately non-adherent. This means they were missing doses, forgetting refills, or stopping meds when they felt better.
• It wasn't usually because they were lazy; it was because the routine was too complex or they were forgetful.

The Top Two "Weight Makers" (Predictors):
The AI found that two things made the backpack feel heaviest and made people more likely to drop it:

1. Polypharmacy (Too Many Pills): The more different medicines a person took, the heavier the burden and the more likely they were to miss a dose. It's like trying to juggle 10 balls instead of 3; eventually, you drop one.
2. Needing Assistance: Patients who needed help from a family member or caregiver to manage their meds felt a heavier burden.
   • The Twist: Interestingly, having help sometimes helped, but if a person needed help and had a complex regimen, the risk of missing doses went up. It suggests that if the system is too broken for a person to handle alone, even a helper might struggle to keep up.

Other Surprising Clues:

• Money Matters: Patients who had to pay for their own meds out of their own pocket felt a heavier burden and were less likely to stick to the plan.
• Education & Work: Surprisingly, people with higher education and those who were unemployed (and thus had more time) were actually better at sticking to their meds. Being busy with a job made it harder to remember to take pills.

4. The "Magic" of the AI (Machine Learning)

The study showed that the AI models (specifically one called Random Forest) were much better at predicting these problems than traditional math.

• Why? Because traditional math assumes everything is a straight line (e.g., "More pills = More burden"). The AI realized that the relationship is a curve. For example, having 5 pills might be fine, but having 6 might suddenly make the burden skyrocket if you also have to pay for them and have no help. The AI caught these "tipping points."

5. What This Means for the Future

The researchers are saying: "We need to lighten the load."

• Simplify the Regimen: Doctors should try to prescribe fewer pills or combine them into one pill if possible.
• Patient-Centered Care: Instead of just handing out a prescription, doctors need to talk to patients about how they will manage it. "Can you afford this?" "Who will help you remember?"
• Deprescribing: This is the art of safely stopping medicines that aren't needed anymore. The study found that many patients want to stop some meds but are afraid to ask. Doctors need to lead these conversations.

The Bottom Line

In Central Nepal, older adults are carrying a heavy "medicine backpack." The study used smart computer programs to figure out that too many pills and needing help to manage them are the biggest reasons people struggle.

The solution isn't just telling people to "try harder." It's about doctors and the healthcare system simplifying the routine, helping with the costs, and making sure the "backpack" is light enough for the patient to carry every day without dropping it.

Technical Summary

1. Problem Statement

Nepal is undergoing a rapid demographic shift toward an aging population, accompanied by a high prevalence of non-communicable chronic diseases (NCDs). This demographic transition leads to increased polypharmacy, multimorbidity, and the use of potentially inappropriate medications (PIMs).

• Core Issue: Older patients face a significant Medication-Related Burden (MRB)—the practical and psychological challenges of managing complex regimens—which often leads to medication non-adherence.
• Knowledge Gap: While the prevalence of these issues is known, the multidimensional, non-linear interactions between socio-demographic, clinical, and medication-related factors driving MRB and adherence in the Nepalese context remain poorly understood. Traditional statistical methods often fail to capture these complex, non-linear relationships.
• Objective: To assess MRB and refill adherence among older ambulatory patients in Central Nepal and utilize Machine Learning (ML) to identify the complex predictive features influencing these outcomes.

2. Methodology

Study Design and Population

• Design: Cross-sectional study.
• Setting: Bharatpur Hospital, a tertiary public hospital in Central Nepal.
• Participants: 390 ambulatory patients aged ≥65 years with at least one chronic condition requiring medication for >1 year.
• Exclusion Criteria: Patients with psychiatric disorders or cognitive impairment.
• Sample Size Calculation: Determined a priori using the pmsampsize R package for multivariable prediction modeling, targeting 372 participants (adjusted to 390 to account for 5% missing data) to limit shrinkage of the apparent $R^2$ to ≤10%.

Data Collection and Instruments

• Outcome Variables:
  1. Medication-Related Burden (MRB): Measured using the Living with Medicines Questionnaire (LMQ-3). This 41-item tool covers 8 domains (e.g., practical difficulties, cost, side effects, autonomy) on a 5-point Likert scale. Total scores range from 41 to 205 (higher = higher burden). A Visual Analog Scale (VAS) was also used.
  2. Medication Adherence: Measured using the Adherence to Refills and Medications Scale (ARMS). A 12-item scale (4-point Likert) assessing refill adherence behaviors. Scores range from 12 to 48 (higher = worse adherence).
• Predictor Features: 17 candidate features were identified, categorized into:
  • Socio-demographic: Age, gender, residence, marital status, occupation, education, ethnicity, economic status.
  • Clinical: Primary diagnosis, multimorbidity, Charlson Comorbidity Index (CCI).
  • Medication-related: Polypharmacy status, formulation type, frequency, need for assistance, out-of-pocket payment, duration of use.
• Translation: Tools were translated into Nepali using forward-backward translation and cognitive debriefing.

Machine Learning Framework

• Software: R (using tidymodels, recipes, and pmsampsize packages).
• Algorithms Evaluated: Six architectures were trained and compared:
  1. Ordinary Least Squares (OLS) - Baseline.
  2. Penalized Linear Regression (LR).
  3. Random Forest (RF).
  4. Extreme Gradient Boosting (XGBoost).
  5. Light Gradient Boosting Machine (LightGBM).
  6. Support Vector Machines (SVM).
• Validation Strategy:
  • Data split: 80% training / 20% hold-out test set (stratified sampling).
  • Hyperparameter tuning: 10x10 repeated cross-validation with ANOVA-based racing.
  • Performance Metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), $R^2$, Mean Absolute Percentage Error (MAPE), and Huber loss.
• Interpretability: SHAP (SHapley Additive exPlanations) values were used to rank feature importance, visualize global interpretability, and analyze feature interactions (dependence plots).

3. Key Results

Descriptive Statistics

• Demographics: 64.9% were "young-old" (65–74), 55.4% female, 68.2% illiterate, and 77.4% had multimorbidity. Polypharmacy affected 70.8% of participants.
• Burden (LMQ-3): Median total score was 110.0 (IQR 14.0), indicating moderate MRB. The highest burden domains were "Practical difficulties" and "Attitudes/Concerns."
• Adherence (ARMS): Median score was 21.0 (IQR 6.0), indicating moderate non-adherence. Common issues included forgetting doses and running out of medication.

Predictive Modeling Performance

• Best Model: Random Forest (RF) emerged as the superior model for both outcomes, outperforming traditional OLS and other ML models.
  • LMQ Prediction: RF achieved $R^2 = 0.540$ (vs. OLS $R^2 = 0.476$) and the lowest RMSE (7.526).
  • ARMS Prediction: RF achieved $R^2 = 0.495$ (vs. OLS $R^2 = 0.365$) and the lowest error rates.
• Implication: The relationship between patient characteristics and medication behaviors is non-linear and complex, which ML captures better than linear regression.

Feature Importance and Drivers (SHAP Analysis)

• Top Predictors for MRB:
  1. Requires Assistance: The single most influential predictor.
  2. Polypharmacy: Strongly associated with higher burden.
  3. Clinical Factors: CCI category and multimorbidity were consistent middle-tier predictors.
  4. Financial: Out-of-pocket payments increased burden.
  • Note: Socio-demographics (age, gender, ethnicity) had the lowest predictive power.
• Top Predictors for Adherence:
  1. Polypharmacy: The strongest predictor of poor adherence.
  2. Requires Assistance: Significant driver of non-adherence.
  3. Financial/Economic: Paying for medications and employment status were key.
  • Counter-intuitive finding: Being unemployed and having post-primary education were associated with better adherence (negative predictors), likely due to more time for management and higher health literacy.

Interaction Effects

• Assistance vs. Complexity: While polypharmacy generally worsens adherence, the risk is partially mitigated when patients receive physical or cognitive assistance. However, patients with simple regimens who still require assistance showed a "peak risk" for poor adherence, suggesting functional/cognitive decline overrides regimen simplicity.
• Financial Stress: The burden of out-of-pocket payments consistently increased MRB and non-adherence, especially when compounded by the need for assistance.

4. Key Contributions

1. First ML Application in Context: This is the first study in South Asia to apply machine learning architectures to predict medication burden and adherence in an older Nepalese population.
2. Superior Predictive Accuracy: Demonstrated that ML (specifically Random Forest) significantly outperforms traditional linear models in this domain, validating the non-linear nature of medication management challenges.
3. Identification of Critical Drivers: Shifted the focus from traditional demographics (age/gender) to functional and regimen-specific factors (Polypharmacy and Need for Assistance) as the primary drivers of burden and non-adherence.
4. Policy-Relevant Insights: Highlighted the specific vulnerability of patients requiring assistance and the impact of out-of-pocket costs, providing data for targeted interventions.

5. Significance and Conclusion

The study concludes that older adults in Central Nepal face a moderate but significant medication burden and non-adherence, driven largely by regimen complexity and dependency on others.

• Clinical Implications: Interventions should prioritize simplified regimens and patient-centered counseling, specifically targeting those with high dependency (needing assistance) and polypharmacy.
• Deprescribing: The findings support the implementation of deprescribing protocols, as over half of patients in Nepal are willing to reduce medications if possible.
• Systemic Changes: The data supports policy-level strategies to reduce out-of-pocket costs and improve pharmacy access for geriatric populations.
• Future Directions: The study advocates for the integration of ML tools in clinical settings to identify high-risk patients early and tailor interventions, moving beyond "one-size-fits-all" approaches to precision geriatric pharmacotherapy.

Limitations: Single-center design (Bharatpur Hospital) limits generalizability; some LMQ-3 domains showed low internal consistency in the Nepali translation, suggesting a need for further cultural validation.