Deciphering Factors Contributing to Cost-Effective Medicine Using Machine Learning

1. Introduction

Over-the-counter (OTC) products are medications proven to be safe and effective for purchase without a prescription from a physician, treating conditions such as pain, coughs, colds, diarrhea, heartburn, and allergies. They are readily available in pharmacies, grocery stores, gas stations, and online platforms. A survey indicates that 93% of U.S. adults prefer using OTC medicines for minor health issues before seeking professional care, with 92% of physicians endorsing their effectiveness and safety [1].

With the rapid development of the Internet, there has been a significant opportunity for collaboration between the medicine retail industry and online platforms. The growing demand for home healthcare and wellness has fueled the expansion of online medicine purchases. Consumers increasingly rely on platforms like Amazon Pharmacy, Health Warehouse, and Optum, projected to account for about 35.26% of total revenue in the OTC pharmaceuticals market by 2024 [2].

Before deciding to buy OTC medicines, consumers go through several stages. These include recognizing a problem or symptom of a disease (problem recognition), finding appropriate information on drug indications (information search), evaluating alternatives (evaluation of alternatives), and finally deciding on the right medication (purchase decision). After purchasing, consumers evaluate whether the medication met their expectations and how satisfied they feel (post-purchase evaluation) [3,4]. Since decisions are largely based on personal experience, they are subject to biases such as age, seriousness of the symptoms, and medication allergies, making the purchasing process complex. Therefore, consumers need to pay more attention to information search and evaluation of alternatives when making medicine purchasing decisions.

Amazon presents a comprehensive selection of medications for conditions like colds, allergies, digestion issues, and pain relief. Each product listing on Amazon provides transparent pricing and user ratings, serving as indicators of perceived value that consider medications' efficacy, quality, and user satisfaction. This information brings convenience and transparency to the online shopping experience. However, consumers are confronted with the daunting task of selecting from a vast array of medicines and their factors to address common conditions. Numerous factors related to OTC medicine impact customer decision-making, such as price, user reviews, efficacy, brand, size, ingredients, and side effects [5,6,7,8,9,10,11].

Price and perceived value are widely recognized as crucial factors influencing consumer decisions [12,13,14,15]. When prices are similar, consumers tend to favor medications perceived to offer higher value. Historically, value was prioritized over price by consumers, who believed higher prices correlated with higher value (“you get what you pay for”). However, recent research suggests a shift in consumer behavior towards seeking cost-effective options. Gao et al. discovered that while price and perceived value are positively associated, higher prices do not always equate to proportionally higher perceived value [16]. Some consumers prioritize value over price, exhibiting lower price sensitivity. For instance, those seeking top-value products are less concerned with price and more willing to pay for perceived quality, while price-sensitive individuals may choose a lower-priced option even if it offers less perceived value than a higher-priced alternative.

Perceived value is defined as the psychological balance consumers strike between expected gains and sacrifices in transactions [17,18,19]. The correlation between value and price is a pivotal area of interest across industries, addressed prominently by Nelson's “Quality-Price Tradeoff” theory [20]. According to this theory, consumers weigh perceived value against price when making purchasing decisions, expecting higher value as prices increase. Creyer and Ross used a value index to show that consumers often opt for lower-priced, higher-value options over higher-priced, higher-quality ones [21]. Similarly, Zeithaml emphasized that product value—what consumers receive relative to what they pay—is critical in consumer decision-making [22]. Yoon et al. noted that shoppers use a value index (value = quality/price) to guide their purchasing decisions [23].

As medication costs rise significantly—some top-selling drugs have seen over 50% increases in costs since 2012 [24], with projections of further annual increases [25]—consumers face heightened pressure to balance value and price when selecting treatments. This trend underscores the importance of choosing cost-effective products that offer competitive prices and high perceived value. To assist consumers in making informed and cost-effective medication purchases, this paper focuses on employing machine learning techniques to identify key factors influencing medication cost-effectiveness. Specifically, we introduce a novel Cost-Effectiveness Rating (CER) indicator derived from a medicine's user rating relative to its price. This CER provides valuable guidance for consumers navigating product choices based on price considerations.

Machine learning and deep learning are essential components of artificial intelligence, widely applied in healthcare, digital retailing, and social media [26,27,28,29]. In this study, our goal is to simplify customer decision-making when purchasing cost-effective medicines. To achieve this, we utilized machine learning models that incorporate various variables such as medication ingredients, brand, manufacturer, and safety warnings extracted from Amazon web crawls. These models predict the Cost-Effectiveness Rating (CER) of medicines.

We employed a range of machine learning algorithms including Decision Tree (DT), Random Forest (RF), XGBoost, Logistic Regression, and Multilayer Perceptron (MLP) for classification. Decision Trees use nodes and branches to classify data based on attribute values [30], while Random Forests combine multiple decision trees to enhance accuracy [31]. XGBoost improves on traditional gradient boosting with optimizations like parallel tree construction and pruning [32]. Linear Discriminant Analysis (LDA) reduces dimensionality and enhances class separability [33], and the K-nearest neighbor (KNN) method identifies nearby data points based on distance metrics [34]. MLP, a neural network with interconnected layers, learns complex data patterns for effective decision-making [35]. Finally, employing techniques such as SHAP values [36] and logistic regression, we explored the impact of each variable and identified key factors influencing medication cost-effectiveness.

By leveraging the insights gleaned from our analysis, our objective is to empower consumers to make informed and cost-effective purchasing decisions. Through the identification of key factors, we aim to guide consumers towards maximizing perceived value while minimizing costs. This research not only benefits consumers but also provides valuable insights for manufacturers and retailers, enabling them to enhance product competitiveness by focusing on features that drive cost-effectiveness. Ultimately, our work endeavors to enhance consumer welfare and optimize market dynamics in the pharmaceutical industry by facilitating prudent decision-making in the realm of medication purchases.

2. Methodology

2.1. Dataset

Our dataset was sourced from Amazon.com, where we conducted web crawling in Python 3.9.7 to gather information on four common types of over-the-counter (OTC) medicine: cold, allergy, digestion, and pain relief. These categories were displayed on the website as “Cold & Flu Medicine,” “Allergy Medicine,” “Antacids,” and “Non-aspirin Pain Relievers,” respectively. Notably, each medicine type on Amazon encompassed various subcategories. For instance, within the cold medicine category, there existed additional subcategories such as “Cough & Sore Throat Medicine.” To ensure robust data volume for our analysis, we selected the most significant subcategories, such as “Cold & Flu Medicine” and “Antacids” for cold and digestion, respectively. Alternatively, we opted for subcategories with greater diversity compared to others within the same medicine type. For example, “Non-aspirin Pain Relievers” was chosen for pain medicine due to its broader range of treatments, despite being the second largest subcategory next to “Joint & Muscle Pain Relief.” Similarly, “Allergy Medicine” was preferred over “Sinus Medicine” for its potential to provide more general insights into allergy-related symptom treatments. Our dataset comprised 916 records for cold, 618 for allergy, 678 for digestion, and 420 for pain. Each record represented an OTC medicine item available for sale on Amazon, and for each medication, we collected various information, as illustrated by the example item in Table 1.

Table 1 demonstrates the detailed information collected for each item through web crawling, exemplified by the DayQuil and NyQuil Combo Pack. This product, designed to provide multi-symptom relief for cold and flu symptoms, including headache, fever, sore throat, and cough, is priced at $22.99 and has received high ratings from customers. With an average rating of 4.8 stars based on 7081 reviews, the majority of reviewers gave it a 5-star rating (86%). Additionally, the product is eligible for Flexible Spending Account (FSA) or Health Savings Account (HSA) reimbursement and comes in a pack of 72 liquicaps. We also collected information about the product's dimensions, brand, manufacturer, ingredients, special features, benefits, and usage instructions. Safety information and warnings ensuring safe usage of the product were also included. The ASIN (Amazon Standard Identification Number) and a direct link to the product on Amazon are provided for reference. Subsequently, using this dataset, we developed the “Cost-Effectiveness Rating (CER)” metric for each item by dividing its average rating by its price. In the case of this example item, the CER is calculated as 0.2088 (4.8/22.99). The rest of the attributes in the dataset serve as input factors to analyze their impact on the CER in this research.

2.2. Data Preprocessing

To build appropriate machine learning models to identify factors impacting the cost-effectiveness rating of each medicine item, we conducted data preprocessing. Initially, we removed items lacking ratings or prices, as we couldn't derive the CER for them. This resulted in a reduction of the dataset by 176 records. Additionally, to ensure the robustness of the derived CERs, we eliminated medicine items with fewer than 100 reviews, further reducing our dataset to 1445 records. Upon analyzing the distribution of CERs for each medicine type, we observed that the dataset exhibited right skewness, with some extreme CER values on the right end, as depicted in Figure 1. Consequently, building models directly predicting raw CERs might pose challenges in achieving accuracy, and the derived important factors from such models may not be reliable enough [37]. Therefore, we engineered a binary target variable using the median CER of each medicine type as a benchmark. If the CER was above the median, it was assigned a value of 1; otherwise, it was assigned a value of 0. Subsequently, we developed separate machine learning models for each medicine type to predict this binary target variable based on the CER and to identify important factors contributing to cost-effective medicine. For each machine learning model constructed, we employed 5-fold cross-validation to ensure the robustness of the model.

2.3. Data Exploration and Feature Engineering

In order to develop effective machine learning models for predicting the binary CERs of each medicine type, we conducted thorough exploration and feature engineering. Below is a comprehensive breakdown of the features created for each of the four medicine types, detailed in Table 2. These features are categorized into eight groups: FSA or HSA eligibility, Size metrics, Brand, Manufacturer, Active Ingredients, Special Effects, Symptom Treats, and Safety Warnings.

2.3.1. FSA or HSA Eligible

This denotes whether the medicine item is eligible for Flexible Spending Account (FSA) or Health Savings Account (HSA) benefits, indicated as Yes or No. The distribution of FSA status across all collected medicine items is depicted in Figure 2.

2.3.2. Size

Under this category, we considered three metrics to gauge the size perspective of each item: counts per pack, weight (in ounces), and dimensions (in inches). Each metric was divided into four quantiles, and then one-hot encoding was utilized to create dummy variables for each quantile. Figure 3 displays the distribution of counts per pack from all collected medicines.

• Counts per Pack: Refers to the number of items per pack.
• Weight: Represents the weight of the item.
• Inches: Indicates the dimensions of the item.

2.3.3. Brand

This refers to the brand of the item. We counted the frequency of each brand by medicine type and then converted major brands into binary columns using one-hot encoding. Figure 4 shows the distribution of the top 10 brands across all collected medicines.

2.3.4. Manufacturer

Similar to the Brand feature, we counted the frequency of each manufacturer by medicine type and created individual binary columns for major manufacturers using one-hot encoding. Figure 5 displays the distribution of the top 10 brands among all collected medicines.

2.3.5. Active Ingredients

This includes the active ingredients of the item. We counted the frequency of each active ingredient by medicine type and then converted major active ingredients into binary columns using one-hot encoding. Figure 6 demonstrates the distribution of the top 10 ingredients across all collected medicines.

2.3.6. Special Effects

These features indicate whether the medicine item possesses specific properties such as fast-acting, long-lasting, maximum strength, non-drowsy, and/or kid-friendly. Keywords reflecting each special effect were identified from columns like 'Product Name,' 'Special Feature,' 'About,' and 'Item Description’. Figure 7 illustrates the distribution of maximum strength, non-drowsy, kid-friendly, and long-lasting special effects among all collected medicine items.

2.3.7. Symptom Treats

This refers to the number of symptom words each medicine item treats. Major symptom words were identified from columns like 'Special Benefit,' 'Special Use,' 'About,' and 'Item Description,' based on their frequencies by medicine type. A variable was created to count how many major symptom words each medicine has. Figure 8 presents the distribution of the top 10 symptom words treated by all collected medicine items.

2.3.8. Safety Warnings

This indicates the number of safety concern words associated with each medicine item. Major safety concerns were identified from the 'Safety Information' column by counting their frequencies. A variable was created to count how many major safeties concern words each medicine item has. Figure 9 showcases the distribution of the top 10 safety concern words among all collected medicine items.

2.4. Machine Learning Modeling and Impact Assessment of Key Factors

In the previous sections, we established a binary CER target variable based on median values for predicting Cost-Effectiveness Ratings. We engineered features across eight categories, including FSA or HSA eligibility, Size metrics, Brand, Manufacturer, Active Ingredients, etc. Subsequently, for each medicine type, we employed eight machine learning models/classifiers to predict binary CER, each validated through 5-fold cross-validation. Using various metrics, we identified the optimal machine learning model with the best hyperparameter set for each medicine type. SHAP values were then utilized to evaluate feature importance and identify key feature categories influencing CER for each medicine type. Furthermore, logistic regression was employed to determine the direction of impact—whether positive (indicating greater cost-effectiveness) or negative (indicating lesser cost-effectiveness)—of key factors.

2.4.1. Machine Learning Models for Predicting CER and Performance Metrics

Using Python 3.9.7, we selected eight machine learning models/classifiers to predict binary CER for each of the four medicine types: Logistic Regression (LR), K-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF), XGBoost (XGB), Linear Discriminative Analysis (LDA), Gaussian Naïve Bayes (Gaussian NB), and Multi-Layer Perceptron Classifier (MLP). These models were chosen based on their previous success in relevant predictions [26,27,28,29,30,31,32,33,34,35] and unique strengths. We evaluated each model through 5-fold cross-validation using the collected medicine dataset and refined them using GridSearchCV (GSCV) from the Scikit-Learn library for hyperparameter tuning. We began with Logistic Regression for its simplicity, then explored the instance-based approach of KNN for its ability to capture non-linear relationships. Subsequently, to address the inefficiency of KNN in high-dimensional datasets with numerous features, we delved into tree-based models such as Decision Tree, Random Forest, and XGBoost. Additionally, to complement our approach, we integrated LDA and Gaussian NB for their probabilistic characteristics. Lastly, we employed MLP, a neural network model, for its flexibility in capturing complex relationships and loose assumptions, despite MLP’s reduced interpretability compared to other models.

To assess the effectiveness of our classifiers, we employed established metrics, including accuracy, Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC), precision, recall, and F1-score. These metrics provide comprehensive insights into the performance of our models. Accuracy measures the proportion of correct predictions, while the ROC-AUC curve visually represents classifier performance by plotting recall against the false positive rate across various thresholds, condensed into a single metric. Precision gauges the reliability of positive classifications, while recall denotes the fraction of actual positives correctly identified, highlighting the impact of false negatives. Lastly, the F1-score offers a balanced assessment of precision and recall.

2.4.2. SHAP Values and Logistic Regression Coefficients for Identifying Factor Impact

Multiple methods are commonly employed to identify key features within machine learning models. One method entails leveraging Feature Importance from Tree-Based Models, such as decision trees, random forests, and gradient boosting machines. These models calculate feature importance using metrics like Gini impurity or entropy, which assess a feature's effectiveness in data splitting and uncertainty reduction. Another approach involves examining Model Coefficients, like those in logistic regression, which indicate the direction and strength of the relationship between features and predicted outcomes. Another approach, Permutation Importance assesses performance decreases when feature values are randomly permuted. However, a particularly powerful method is SHAP (SHapley Additive exPlanations), which interprets machine learning model outputs by attributing predictions to each feature's contribution. By employing concepts from cooperative game theory [36], SHAP values provide insights into each feature's impact on individual predictions. Unlike other methods, SHAP values offer a unified framework for interpreting complex machine learning models, regardless of whether they are logistic regression, tree-based models, or neural networks, making them highly versatile. Additionally, SHAP values inherently account for multicollinearity among features by considering their joint contributions to model predictions [38]. In contrast, tree model feature importance, model coefficients, or permutation importance may overlook multicollinearity issues or interactions between features, potentially leading to biased importance scores.

By analyzing mean absolute SHAP values, we determined the magnitude of each factor's impact. Further understanding the direction of impact, logistic regression models were developed to analyze the sign of the coefficient (whether positive or negative) for each factor concerning each medicine type. While SHAP values can also provide insights into the direction of impact, logistic regression offers a straightforward and intuitive interpretation [36,39]. A positive coefficient for a particular factor indicates that an increase in the feature value leads to a higher likelihood of the positive class, thus being more cost-effective. Conversely, a negative coefficient suggests that an increase in the feature value corresponds to a lower likelihood of the positive class of CER. The direct relationship between the sign of the coefficient and the prediction outcome facilitates an easy interpretation of the directional impact of factors on cost effectiveness rating. Moreover, compared to SHAP, logistic regression coefficients offer a simpler approach for customers to grasp the directional influence of factors, aiding in making more cost-effective purchasing decisions.

3. Results

3.1. Machine Learning Classifiers for CER Across Medicine Types

Table 3, Table 4, Table 5 and Table 6 present an evaluation of eight machine learning classifiers across the four medicine types (cold/allergy/digestion/pain relief), utilizing a 5-fold cross-validation methodology to predict binary Cost-Effectiveness Ratings (CERs). Results are reported as average values with standard deviations. While accuracy and F1 metrics are essential indicators of predictive performance, we primarily emphasize the ROC-AUC metric due to its threshold independence and ability to assess the model's ranking capabilities, crucial for correctly identifying true positives, especially in the context of high-cost-effective medications.

Upon through examination of the results for each medicine type, the choice of the most suitable model varies depending on the specific medication under consideration. For cold medicine, the Random Forest (RF) model emerges as the most effective choice, achieving the highest ROC-AUC of 0.7428 ± 0.0863 among all models, indicating its superior ability to discern between high and low CERs. For allergy medicine, despite the simplicity of the Logistic Regression (LR) model, it demonstrates robust performance with a ROC-AUC of 0.7548 ± 0.045, outperforming alternative models such as Linear Discriminant Analysis (LDA). Furthermore, LR exhibits higher average accuracy (0.6793 vs. 0.6480) and average F1 score (0.6849 vs. 0.6394) compared to LDA, justifying its preference. In the case of digestion medicine, Random Forest (RF) once again showcases its effectiveness with a commendable ROC-AUC of 0.7081 ± 0.071, surpassing Logistic Regression and XGBoost in predictive capability, as supported by higher accuracy and F1 scores. Lastly, for pain relief medicine, Random Forest (RF) stands out with the highest ROC-AUC of 0.8022 ± 0.050, underscoring its robust performance and versatility in handling diverse features.

In summary, while Random Forest (RF) consistently demonstrates commendable performance across different medicine types, the optimal model choice varies based on the unique characteristics and complexities of each medication's dataset. Consequently, for cold medicine, allergy medicine, digestion medicine, and pain relief medicine, the preferred models are Random Forest (RF), Logistic Regression (LR), Random Forest (RF), and Random Forest (RF), respectively. Subsequent sections will delve into the analysis of important features or input factors using the identified best model for each medicine type, assessing their impact on cost-effectiveness ratings.

3.2. Key Feature Categories Influencing CER Across Medicine Types

Figure 10, Figure 11, Figure 12, and Figure 13 provide insights into the primary factors influencing Cost-Effectiveness Ratings (CER) across cold, allergy, digestion, and pain relief medicines. These insights are derived from SHAP values calculated using the best model identified for each medicine type. By examining the top five factors' feature categories in each plot, we discerned the most impactful feature categories for each medicine type from the eight feature categories included in this research (FSA or HSA eligibility, Size metrics, Brand, Manufacturer, Active Ingredients, Special Effects, Symptom Treats, and Safety Warnings).

In Figure 10, we discerned the key feature categories influencing the CERs of cold medicine using the Random Forest model:

• FSA or HSA Eligibility: Signifying the potential for consumers to utilize pre-tax funds for medication purchases, which may be viewed as more cost-effective.
• Symptom Treats: The number of symptoms treated emerges as a significant contributor to CER, underscoring the importance of efficacy considerations.
• Safety Warnings: The presence of safety warnings also significantly contributes to cost-effectiveness ratings, emphasizing the importance of safety considerations.
• Size Metrics: Both lower and higher quantiles of inches play a significant role in influencing CER, suggesting that the physical dimensions of the medication packaging impact its cost-effectiveness.

In Figure 11, using the Logistic Regression model, we examine the factors influencing the CERs of allergy medicine and identified key feature categories:

• Size Metrics: Particularly, smaller-sized packaging or lighter weight contribute to actual cost-effectiveness.
• Manufacturer Influence: Specific manufacturers like Johnson & Johnson, Bayer, Sanofi, Major, and Perrigo exert notable influence, indicating that brand reputation and trustworthiness may affect consumer ratings when adjusting the cost
• Special Effects: Attributes like being kid-friendly influence CER, enhancing safety perceptions and influencing actual cost-effectiveness.
• Symptom Treats: Similarly to cold medicine, the medication's ability to address a broader range of symptoms impacts CER.

Figure 12 showcases the factors influencing the CERs of digestion medicine using the Random Forest model, where we identified key feature categories:

• FSA or HSA Eligibility: Similarly to cold medicine, suggesting the potential for pre-tax fund utilization to be more cost-effective.
• Size Metrics: Similar to allergy medicine, smaller-sized packaging or lighter weight particularly affect actual cost-effectiveness.
• Symptom Treats: Similarly to cold and allergy medicine, the medication's effectiveness in treating a range of symptoms influences cost-adjusted ratings.
• Active Ingredients: Specific ingredients like Calcium, Famotidine, and Magnesium influence perceived cost-effectiveness.

Figure 13 uncovers the factors impacting the CERs of pain relief medicine using the Random Forest model and identified key feature categories:

• Size Metrics: Both lower and higher quantiles of inches impact packaging dimensions.
• FSA or HSA Eligibility: Signifying potential pre-tax fund usage to be more cost-effective.

These insights offer a comprehensive view of the feature categories driving cost-effectiveness across different medicine types, with further detailed analyses of individual factors from these categories presented in the Discussion section.

4. Discussion

Building upon the insights gleaned from the SHAP plots presented in the previous section, which evaluated the relative importance of various factors and identified key feature categories, we proceeded to develop logistic regression models for each of the four medicine types. These models enable us to distinguish between positive effects (indicating higher cost-effectiveness) and negative effects (indicating lower cost-effectiveness) on Cost-Effectiveness Ratings based on the sign of each coefficient. Figure 14 showcase the directional impact of factors, as inferred from logistic regression coefficients, with highlighted factors across cold, allergy, digestion, and pain relief medicine prominently featured.

In Figure 14, we examined the directional impact of key factors for cold medicine. Both 'FSA or HSA Eligible' and 'Symptom Treats Count' showed positive impacts, indicating that medicines eligible for pre-tax funds and those treating more symptoms tend to be more cost-effective. Surprisingly, 'Safety Warning Count' also positively influenced CER, suggesting that medicines with safety warnings might offer better cost-effectiveness compared to those without. When comparing medicines with and without safety warnings, we found that those with warnings not only had a lower average price ($12.95 vs. $19.08) but also received higher average ratings (4.68 vs. 4.57). Further analysis revealed that medicines with safety warnings more frequently contained active ingredients such as dextromethorphan, acetaminophen, and phenylephrine (as shown in Table 7), clinically proven to be effective in treating cold symptoms [40,41,42]. Moreover, Figure 4-1 highlights phenylephrine and acetaminophen as top factors positively impacting CER, indicating that the inclusion of such ingredients contributes to higher ratings for medicines with safety warnings when the price is the same. Therefore, we do not discourage the purchase of cold medicines with safety warnings. They offer cost-effectiveness due to their lower average price and the inclusion of effective ingredients such as phenylephrine and acetaminophen, resulting in higher ratings. However, individuals should consider their allergies before opting for these medicines. Additionally, smaller packaging positively impacts cost-effectiveness, while larger packaging has a negative effect.

In Figure 15, we delved into allergy medicine and uncovered insights into the directional impact of key factors. We found that factors like smaller-sized packaging and lighter weight held positive coefficients, affirming their role in improving cost-effectiveness. Moreover, allergy medications featuring kid-friendly special effects demonstrated heightened cost-effectiveness, as indicated by their positive coefficient. Additionally, akin to cold medicine, allergy remedies addressing a broader array of allergy symptoms generally received higher ratings at comparable prices, thus bolstering cost-effectiveness. When scrutinizing manufacturers, we observed negative coefficients for Johnson & Johnson, Bayer, and Sanofi, while Major and Perrigo exhibited positive coefficients. However, a closer examination, as Table 8 shows, of the average price, rating, and Cost-Effectiveness Ratings (CER) by these manufacturers revealed conflicting outcomes. Despite Perrigo and Major achieving slightly higher ratings, their elevated average prices outweighed the benefits, resulting in lower average CER values, indicating reduced cost-effectiveness. This suggests that interactions may have existed between manufacturer and other feature categories such as brand, ingredients, and safety warnings, collectively influencing cost-effectiveness ratings and thus twisting interpretations of manufacturer logistic coefficient [43,44]. Consequently, relying on manufacturer-based decisions may have lacked robustness in guiding consumers towards cost-effective allergy medicine purchases. Therefore, in assessing allergy medicine, we primarily focused on other key feature categories identified for CER, particularly size metrics such as smaller size or lighter weight, special effects—especially those appealing to children—and symptom coverage, particularly medicines capable of addressing a broader range of symptoms.

In Figure 16, we analyzed the directional impact of key factors in digestion medicine. Similar to cold medicine, being FSA or HSA eligible proved to be more cost-effective, as confirmed by its positive coefficient. Likewise, akin to allergy medicine, smaller-sized packaging or lighter weight also demonstrated increased cost-effectiveness, as indicated by their positive coefficient. Furthermore, akin to both cold and allergy medicine, addressing a broad range of digestion symptoms was shown to be more cost-effective. Regarding active ingredients, Calcium, Famotidine, and Magnesium all exhibited positive coefficients, indicating increased cost-effectiveness. Based on the collected data, top digestion brands containing calcium included Prelief (with 85.71% labeling calcium as an active ingredient), Rolaids, Tums, and Mylanta. For famotidine, Pepcid stood out, with 57.14% labeling famotidine as an active ingredient. Finally, top digestion brands for magnesium were Rolaids (with 71.43% labeling magnesium as an active ingredient) and Mylanta.

In Figure 17, analyzing pain relief medicine, smaller-sized packaging positively impacts CER, while larger-sized packaging negatively impacts CER. Items eligible for FSA or HSA are more cost-effective.

6. Conclusions

This study used machine learning to identify key factors influencing the cost-effectiveness of over-the-counter (OTC) medications. The analysis revealed that FSA/HSA eligibility, symptom treatment range, active ingredients, special effects, safety warnings, and packaging size significantly impact cost-effectiveness across cold, allergy, digestion, and pain relief medications. Medications eligible for FSA or HSA funds, those treating a broader range of symptoms, and those with smaller packaging are generally perceived as more cost-effective. For cold medicines, the presence of safety warnings does not compromise cost-effectiveness due to their lower average price and the inclusion of effective ingredients such as phenylephrine and acetaminophen. Allergy medications featuring kid-friendly special effects demonstrated heightened cost-effectiveness. Active ingredients like calcium, famotidine, and magnesium notably boost the cost-effectiveness of digestion medicines. Consumers can use these insights to make more informed choices, ensuring they get high-quality treatments at optimal prices. For manufacturers and retailers, emphasizing these key factors can improve product appeal and competitiveness. Overall, leveraging machine learning to understand cost-effectiveness helps improve decision-making for consumers, manufacturers, and retailers in the pharmaceutical industry.