Explainable machine learning for revisiting reported Irritable Bowel Syndrome correlates in a student cohort

By reanalyzing a dataset of Bangladeshi students with Explainable Boosting Machines and addressing data quality issues, this study identifies psychological distress, elevated BMI, and academic dissatisfaction as key predictors of Irritable Bowel Syndrome while revealing non-linear patterns and weaker gender associations that diverge from previous logistic regression findings.

The Gist

Imagine you have a recipe for a very popular cake (Irritable Bowel Syndrome, or IBS) that was written by a previous chef (Hasan et al.). That recipe claimed to know exactly which ingredients cause the cake to rise or fall. However, when two new chefs (Ramirez-Lopez and Kang) looked at the original recipe book, they noticed something strange: the book had some pages where a man was listed as having a "period," and some measurements of height and weight were physically impossible (like a person weighing 300 kg or being 4 meters tall).

So, these new chefs decided to audit the recipe book, throw out the impossible pages, and then try to bake the cake again using a super-smart, modern kitchen assistant (Explainable Boosting Machines, or EBMs) instead of the old, simple calculator the first chef used.

Here is what they found, broken down into simple concepts:

1. The "Garbage In, Garbage Out" Problem

Before they even started baking, they had to clean up the data. They found 44 records that didn't make sense (like the men with periods). If you try to bake a cake with rotten eggs, the cake will taste bad, no matter how good your oven is. By removing these "rotten eggs," they reduced their sample from 550 students to 506, but the results became much more reliable.

2. The Old Calculator vs. The Smart Assistant

• The Old Way (Logistic Regression): The original study used a simple linear tool. Think of this like a ruler that only measures straight lines. It assumes that if you add more of an ingredient (like stress), the result (IBS) goes up in a perfectly straight line. It also couldn't see how ingredients might mix together to create a surprise effect.
• The New Way (Explainable Boosting Machines): The new chefs used an AI tool that is like a detective who can see curves and hidden connections. It doesn't just look at ingredients one by one; it understands that:
  • Sometimes, adding a little bit of an ingredient helps, but too much hurts (non-linear).
  • Sometimes, two ingredients only cause a problem when they are mixed together (interactions).
  • Crucially, this AI is "explainable," meaning it can tell you why it made a decision, just like a human chef explaining their process.

3. The Big Surprises (What Changed?)

When they compared the new "Smart Assistant" results with the old "Ruler" results, they found some major differences:

• The "Stress" Factor (The Heavyweight Champion): Both methods agreed on one thing: Psychological distress (anxiety and stress) is the biggest culprit. It's like the main ingredient in the cake. If you have high stress, the risk of IBS goes up significantly. This confirms that the gut and the brain are tightly linked (the "gut-brain axis").
• The "Gender" Myth: The old study said being female was a major risk factor. The new study said, "Actually, not so much." Why? Because when you account for stress and body weight, being female doesn't add much extra risk on its own. It's like saying "being tall" causes a car crash; maybe tall people drive more, but the real cause is the driving behavior, not the height.
• The "Exercise" Paradox: The old study thought more exercise was always good. The new study found a curved relationship.
  • Analogy: Think of exercise like a car engine. A little bit of driving is fine. A moderate amount is great. But if you redline the engine and drive at 200 mph for hours (very high intensity), you might break something. The study found that students who exercised too much (over 60 minutes a day) actually had a higher risk of IBS, especially if they were also overweight.
• The "BMI" Twist: The old study worried about being undernourished (too thin). The new study found that being overweight (high BMI) was a much stronger predictor. It's like a heavy backpack; the heavier it gets, the more it strains your back (your gut).
• The "Breakfast" Surprise: The old study didn't find a link to skipping breakfast. The new study found that skipping breakfast was actually linked to a lower risk of IBS in this specific group, but only when looking at how it interacted with body weight. It's a weird finding that suggests the relationship is complex, not a simple "breakfast is good" rule.

4. The "Secret Sauce" (Interactions)

The smart AI found that variables don't just act alone; they dance together.

• Example: Being overweight plus exercising intensely creates a specific risk that neither factor has on its own.
• Example: Not choosing your own university major (feeling forced into a path) was a strong predictor of IBS, likely because it causes stress.

The Takeaway

This paper is a lesson in re-checking the work.

1. Data Cleaning Matters: If you don't fix the "men with periods" errors, your conclusions are wrong.
2. Complexity Matters: Life isn't a straight line. People's bodies react in curves and combinations. Simple math tools miss these nuances.
3. AI Can Be Honest: You can use powerful machine learning to find complex patterns without losing the ability to understand why the model made those predictions.

In short: The original study gave a good first draft, but by cleaning the data and using a smarter, more flexible tool, the new study revealed that stress, body weight, and the intensity of exercise are the real drivers of IBS in students, and the story is much more complicated (and interesting) than we first thought.

Technical Summary

1. Problem Statement

The study addresses the limitations of a previous analysis by Hasan et al. [1] regarding the correlates of Irritable Bowel Syndrome (IBS) among university students in Dhaka, Bangladesh. The original study relied on conventional logistic regression, which the authors argue suffered from four key methodological flaws:

1. Data Anomalies: The original dataset contained biologically implausible records (e.g., males reporting menstrual symptoms) that were not filtered, potentially introducing misclassification bias.
2. Lack of Interaction Modeling: The original analysis focused on main effects, ignoring potential synergistic or conditional interactions between variables (e.g., how BMI might moderate the effect of physical activity).
3. Discretization of Continuous Variables: Continuous predictors (like psychological scores and BMI) were categorized, reducing statistical power and masking non-linear relationships.
4. Linearity Assumption: The original model assumed linear relationships between predictors and IBS risk, which may yield biased estimates if the true associations are non-linear.

The goal was to re-analyze the dataset using a more robust framework that could capture non-linearities and interactions while maintaining interpretability, and to perform a rigorous data audit to ensure validity.

2. Methodology

Data Source and Preprocessing

• Dataset: Secondary analysis of a publicly available cross-sectional survey of 550 undergraduate students in Dhaka (Hasan et al. [1]).
• Data Audit & Cleaning: The authors performed a validation audit, removing 44 observations (reducing the sample to $N=506$).
  • Plausibility Filters: Removed records with unrealistic anthropometrics (height >300cm, weight <0kg or >200kg, BMI <10 or >40).
  • Integrity Check: Excluded 40 male respondents who reported menstrual-related abdominal discomfort, a biological inconsistency absent from the original analysis.
• Outcome Variable: IBS status was reconstructed using the Rome III diagnostic criteria from the raw questionnaire responses.
• Predictors: Included sociodemographics, academic factors, lifestyle behaviors (smoking, physical activity, diet, sleep), anthropometrics (BMI), and continuous psychological scores (DASS-21: Depression, Anxiety, Stress).

Modeling Framework: Explainable Boosting Machines (EBM)

• Algorithm: The study employed Explainable Boosting Machines (EBMs), a class of Generalized Additive Models (GAMs) trained via boosting.
• Rationale: EBMs were chosen to balance high predictive performance with interpretability. They can:
  • Capture complex non-linear relationships without pre-specifying functional forms.
  • Model pairwise interactions between variables.
  • Provide additive shape functions that visualize the marginal contribution of each predictor.
• Implementation:
  • Implemented using the ebm package in R.
  • Class Imbalance: Addressed via random oversampling of the minority class (IBS positive) within cross-validation splits.
  • Constraints: Monotonic constraints were applied to psychological scores based on domain hypotheses (higher distress = higher risk).
  • Validation: Evaluated across 100 stratified train-test splits (75/25).

3. Key Results

Model Performance

• The EBM model achieved a mean AUC of 0.852 (SD = 0.055), significantly outperforming the original study's logistic regression (AUC = 0.7486).
• Sensitivity was 0.91, indicating strong ability to correctly identify IBS cases.

Predictor Importance and Associations

• Top Predictors: Psychological distress (specifically anxiety and stress), elevated BMI, and dissatisfaction with one's academic major were the strongest predictors.
• Non-Linear Effects:
  • BMI: Risk remained relatively flat at lower BMI levels but increased sharply for individuals with a BMI >30 (obese range).
  • Physical Activity: Showed a non-linear "U-shaped" or threshold effect. Low-to-moderate activity reduced risk, but high intensity (>60 mins/day) was associated with increased IBS risk.
  • Breakfast Skipping: Contrary to some expectations, skipping breakfast was associated with a lower estimated risk of IBS, particularly in interaction with high BMI.
  • Fruit Consumption: Only infrequent consumption (monthly) was linked to higher risk; daily/weekly intake showed negligible effects.
• Interaction Effects:
  • The model identified significant interactions between BMI and other variables (physical activity, sleep, breakfast habits). For instance, the risk associated with high physical activity was amplified in obese individuals.
• Gender: Unlike the original study, which found gender to be a significant predictor, the EBM model ranked gender as having low predictive importance. The authors suggest gender's effect is likely mediated through psychological and metabolic pathways rather than being an independent risk factor.

4. Key Contributions

1. Data Quality Verification: Demonstrated that secondary analysis of public datasets requires rigorous auditing. The removal of biologically impossible records (e.g., males with menstrual symptoms) was critical for valid inference.
2. Methodological Advancement: Showed that Explainable Boosting Machines can outperform traditional logistic regression in epidemiological settings by capturing non-linearities and interactions without sacrificing interpretability.
3. Revised Clinical Insights:
   • BMI as a Moderator: Highlighted that BMI is not just a linear risk factor but a critical moderator that changes the risk profile of lifestyle factors like physical activity and diet.
   • Nuanced Lifestyle Factors: Challenged the linear assumption that "more exercise is always better" for IBS, identifying a risk threshold at high intensities.
   • Psychological Primacy: Confirmed that psychological distress (anxiety/stress) is the most robust correlate of IBS in this population, consistent across different modeling approaches.

5. Significance

This study underscores the importance of model selection and data validation in epidemiological research. It illustrates that relying solely on linear regression with discretized variables can lead to:

• Underestimation of complex risk patterns (e.g., the specific risk of high-intensity exercise in obese individuals).
• Misattribution of risk factors (e.g., overestimating the independent effect of gender).

By using an interpretable machine learning framework, the authors provide a transparent alternative to "black box" models, allowing domain experts to validate mechanistic hypotheses (e.g., the gut-brain axis) while uncovering complex, non-additive relationships. The findings suggest that interventions for IBS in student populations should prioritize psychological support and consider the nuanced, non-linear impacts of lifestyle factors like exercise intensity and BMI.