Machine Learning-Based Prediction of Dengue Infection Using Clinical and Hematological Data from Bangladesh: A Comparative Analysis of Benchmark and Extended Prediction Models

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Machine Learning-Based Prediction of Dengue Infection Using Clinical and Hematological Data from Bangladesh".

Jane: The gist The study compares Logistic Regression, Random Forest, and XGBoost to predict dengue from clinical and hematological data using two Bangladeshi hospital datasets (2024, n=1,511; 2025,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So, we're looking at this paper by Neesha Naimeen and Victor Miranda Soberanis called "Machine Learning-Based Prediction of Dengue Infection Using Clinical and Hematological Data from Bangladesh: A Comparative Analysis of Benchmark and Extended Prediction Models". Jane It’s about using machine learning to predict dengue fever using clinical and blood test data, specifically from two different Bangladeshi hospital datasets from two thousand twenty-four and two thousand twenty-five <ref:2610.14813#pg1>. Tom The authors are focusing on comparing three specific models: Logistic Regression, Random Forest, and XGBoost to see which one does the best job. Jane What this means for us is that dengue diagnosis in Bangladesh is getting smarter by using these algorithms on real patient data instead of just relying on manual checks or symptoms alone. Tom It’s a comparative analysis, so they aren't just picking one model and calling it the winner; they are setting up a test to see how different machine learning techniques handle this specific medical problem. Jane And the big implication is that we can start thinking about non-invasive ways to predict when dengue might be hitting hard in these regions before symptoms fully overlap with other illnesses. Tom Exactly, it’s not just about building one good model; it’s about comparing the tools to see which approach holds up best across different types of data and time periods.

Lu: From a representation learning side, I'm interested in how they handled the feature engineering because that dictates what these models can actually learn. Jane Right, so they started with an eight common predictors for their benchmark model, which were kept because they showed consistency across both the two thousand twenty-four and two thousand twenty-five datasets <ref:2610.14813#pg1>. Tom That’s right, like Age, Sex, RBC count, HCT percent, Neutrophils percent and Lymphocytes percent. Lu And then they built an extended model for the two thousand twenty-four dataset that added ten more clinical and hematological predictors to see if those extra features actually help the prediction accuracy <ref:2610.14813#pg1>. Jane So they are testing the value of adding new information versus sticking to a simpler, standard set of data points. Tom Which brings us into the core comparison: how these different algorithms perform when you give them different amounts of input data.

Meng: From a production standpoint, I'm always wondering what happens when you try to run models on real-world traffic versus this controlled setting described here. Jane That’s a fair question, Meng; the paper uses two specific datasets, one from two thousand twenty-four with fifteen hundred and eleven patients and one from two thousand twenty-five with twelve hundred and eighty-three patients <ref:2610.14813#pg1>. Tom The comparison between these two years is important because the baseline difficulty of predicting dengue might change depending on what kind of data we get. Meng So they have to be careful about how they generalize this performance gap when we take it out into a hospital setting that doesn't match these specific collection methods.

Lalam: I see this paper as an attempt to systematically evaluate the utility of combining different algorithms, which is valuable because just picking one algorithm and claiming it’s the best without comparison is risky. Jane Right, so they are looking at Logistic Regression, Random Forest, and XGBoost simultaneously to get a holistic view of performance. Lu And I'm particularly interested in how the results from different tree-based models like Random Forest and XGBoost compare when looking at specific metrics like the F1-score or ROC-AUC.

Tom: Well, they found some interesting things there; for instance, in two thousand twenty-four using that benchmark model, Random Forest achieved a very high F1-score of zero point eight one six and an accuracy of zero point seven two three <ref:2610.14813#pg1>. Jane That sounds solid; so when you look at the results for the two thousand twenty-five dataset, XGBoost actually achieved a higher F1-score of zero point nine one nine and an accuracy of zero point eight seven five <ref:2610.14813#pg1>. Meng So, it seems the best performer changes depending on which year's data we’re looking at; that’s something I’d worry about when deploying models because you need stability.

Lu: And then they also noted a specific finding where Random Forest achieved the best negative-class recall among the tree-based models in two thousand twenty-five <ref:2610.14813#pg1>. Jane That suggests that for predicting cases that *aren't* dengue, Random Forest was actually better than XGBoost in that year, which is an interesting nuance. Tom That’s a specific detail; it shows that different models aren't always equally good across all prediction outcomes.

The paper's summary: Jane: So, let's unpack what the paper actually summarizes about the entire study of "Machine Learning-Based Prediction of Dengue Infection Using Clinical and Hematological Data from Bangladesh: A Comparative Analysis of Benchmark and Extended Prediction Models". Tom Basically, the summary is that they took two datasets, two thousand twenty-four with fifteen hundred and eleven patients, and two thousand twenty-five with twelve hundred and eighty-three patients <ref:2610.14813#pg1>. Jane They used machine learning algorithms to predict dengue infection using clinical and hematological data from these two sources. Tom The main point they are making is a comparison between a simple benchmark model that uses eight common predictors for both years, versus an extended model for two thousand twenty-four that adds ten more specific predictors <ref:2610.14813#pg1>. Jane And the summary points out that while the performance gap between the two years was substantial, it wasn't primarily because of class imbalance, which is what they used SMOTE to correct. Tom They also found that when you look at individual predictor importance, Platelet Count was clearly the most important feature in the two thousand twenty-four benchmark model <ref:2610.14813#pg1>. Jane And they also highlighted that RDW-CV was identified as the most consistently influential added predictor across all three algorithms when they included it in their extended set.

Meng: So, what this means for deployment is that if you're building a system, you need to know which features are truly robust and which ones might be noise. Jane Right; so they found Platelet Count was the top feature in two thousand twenty-four but then they noticed its individual predictive power dropped significantly from zero point five two eight in two thousand twenty-four down to zero point eight seven zero in two thousand twenty-five when looking at the individual ROC-AUC <ref:2610.14813#pg1>. Tom That drop is significant; it suggests that just relying on that one feature might not be enough for a stable model across different data sources.

Lu: And the authors are pointing out a real limitation here, which is that they aren't doing enough to evaluate how well these models generalize beyond the single dataset they were built on. Jane That’s true; they only test performance on these specific Bangladeshi hospital datasets, and there isn't much external validation mentioned. Lu And the paper suggests that this lack of external testing is a major gap in the field, which is something I think we should be watching closely as machine learning moves into clinical settings.

Lalam: I agree with Lu; they are showing that simply getting high accuracy on one set doesn't guarantee it will work elsewhere. Jane So, to summarize what we just covered about the summary, it boils down to comparing three algorithms across two years of data, looking at the impact of adding those extra predictors. Tom And the key takeaway is that while they found a strong performance difference between two thousand twenty-four and two thousand twenty-five that difference seems tied more to how much individual predictors are actually working than just having an imbalance problem <ref:2610.14813#pg1>.

Jane: So, so we're seeing this paper as a comparison of model performance on specific datasets, focusing on feature importance like Platelet Count and RDW-CV across the benchmark versus extended models in the "Machine Learning-Based Prediction of Dengue Infection Using Clinical and Hematological Data from Bangladesh: A Comparative Analysis of Benchmark and Extended Prediction Models".

Tom: And that sets us up perfectly for what they suggest next about how to improve these models.

The paper's improvements: Jane: Now let's talk about the suggested improvements in this paper, because researchers always have ideas on how to make things better. Tom They suggested that adding those ten additional clinical and hematological predictors for the two thousand twenty-four dataset into the extended model did indeed improve prediction performance when compared to their benchmark model <ref:2610.14813#pg1>. Jane So, they are suggesting that incorporating more specific features can definitely boost what the models can achieve within a single year’s data context. Meng From an engineering perspective, adding features means more maintenance; if you add ten new ones and one of them starts causing instability at scale, you have to manage that complexity. Jane That’s true; it increases the complexity of the model structure, which is something we have to keep in mind when we think about deployment. Lu And they also identified RDW-CV as a key area for future work because it was consistently influential in their extended set, suggesting that exploring other variables with similar consistency is a good path forward.

Tom: So, the paper suggests focusing on the importance of predictors like RDW-CV because it was stable across the three different algorithms they tested. Jane And I think that’s a practical suggestion for anyone trying to build a model—focus on features that show up consistently across different types of training methods. Tom Exactly, stability is what we need in a feature set, not just chasing the highest number on one specific run; it means the same thing when you test it against Logistic Regression versus XGBoost.

Meng: And I also hear a suggestion about data quality auditing because they pointed out that the year-to-year performance gap might be due to things like measurement resolution differences. Jane That’s a good point; if the data quality is different between two thousand twenty-four and two thousand twenty-five you need to check for that before blaming the model itself <ref:2610.14813#pg1>. Lu So they are pushing for a formal data quality audit module as a way to diagnose whether performance drops are due to clinical variables or just differences in how the data was recorded.

Lalam: I think this moves beyond just tweaking the algorithm and starts looking at the source of the problem, which is much more valuable for real-world application. Jane That’s a big shift—moving from just tuning parameters to questioning the input itself. Tom So, if we look at this paper as a whole, it's not just about getting a prediction score; it’s about being careful about the data you feed into that system.

Jane: Exactly; the core improvement they suggest is that we need to be more rigorous in checking the input variables themselves before we trust any model output.

Conclusion: Tom: So, to wrap up our discussion on "Machine Learning-Based Prediction of Dengue Infection Using Clinical and Hematological Data from Bangladesh: A Comparative Analysis of Benchmark and Extended Prediction Models". Jane We’ve seen that this paper compared Logistic Regression, Random Forest, and XGBoost across two years of data, finding different performance peaks depending on the year. Tom The main thing to remember is that while adding predictors helped improve prediction performance in controlled comparisons, the authors are suggesting we need to look closely at why individual predictors like Platelet Count lose their power when moving from one dataset to another. Jane And they stress that for real-world use, you need a stable feature set and a rigorous way to check if the data quality itself is causing performance shifts. Tom So, the implication is that we can use these machine learning tools for early dengue detection, but only if we treat the input data with a high degree of scrutiny before trusting any output they give us.

Jane: It’s a study that shows how important it is to look at both the numbers and what those numbers actually mean for the people getting diagnoses in Bangladesh. Lu And I think this paper sets a good direction for future research by pushing for better generalization testing outside these specific datasets. Meng From an engineering standpoint, we need to consider how adding features affects long-term stability before we commit to a final deployment pipeline. Lalam And I think the finding about RDW-CV being consistently influential is a solid anchor point for what should be tested next in this area.

Tom: So, that’s our take on "Machine Learning-Based Prediction of Dengue Infection Using Clinical and Hematological Data from Bangladesh: A Comparative Analysis of Benchmark and Extended Prediction Models". Jane Thanks for tuning in to this discussion.