An Interpretable AI Framework for Multiclass Classification of Thalassemia Using Combined CBC and HPLC Biomarkers

Thalassemia is a hereditary blood disorder that affects millions of people around the world, particularly in South Asia, Southeast Asia, and the Mediterranean. It occurs when the body cannot produce enough of the proteins that make up hemoglobin, the substance in red blood cells that carries oxygen. Without enough of these proteins, the body struggles to make healthy red blood cells, leading to anemia and a host of other health complications. For decades, doctors have diagnosed this condition by examining blood samples under a microscope and using specialized machines to measure the different types of hemoglobin present. While accurate, this process is expensive, time-consuming, and relies heavily on the skill of the individual doctor reading the results. As a result, it is difficult to screen large populations quickly, especially in regions where the disease is most common.

In recent years, scientists have turned to artificial intelligence to help automate this process. The idea is to teach a computer to recognize the patterns in blood test results that indicate thalassemia, much like a seasoned doctor does, but with the speed and consistency of a machine. However, creating such a system is not simple. Blood tests produce a vast amount of data, and the patterns for different types of the disease often overlap. Furthermore, in any large group of people being screened, the vast majority are healthy, while only a tiny fraction actually have the disease or carry the genetic trait. This creates a difficult situation for computer programs, which tend to ignore the rare cases and simply guess that everyone is healthy because that is what happens most of the time.

A team of researchers from SR University and Jyothishmathi Institute of Technology & Science in India has developed a new framework to solve these specific problems. They created a computer system designed to look at two different types of blood test results simultaneously: a standard complete blood count, which measures the size and number of red blood cells, and a more detailed test called high-performance liquid chromatography, which separates and measures the specific fractions of hemoglobin. By combining these two sources of information, the researchers aimed to build a model that could not only detect the disease but also distinguish between three distinct groups: people who are healthy, people who carry the genetic trait (and could pass it to their children), and people who have the active, severe form of the disease.

The researchers began with a massive dataset containing records from over 13,000 patients. This collection included detailed measurements of nine key blood markers, along with demographic information like age and gender. The data was heavily skewed, with healthy individuals making up more than 90 percent of the records, while the severe disease cases accounted for less than half a percent. To handle this imbalance, the team first cleaned the data, fixing errors in how the information was recorded and converting age descriptions into simple numbers. They then grouped the thirteen different diagnostic labels found in the original records into the three broader categories needed for their system. Crucially, they used a technique to artificially create more examples of the rare disease cases within the training data, ensuring the computer learned to recognize them just as well as the common healthy cases.

To find the best way to classify these patients, the team tested four different types of advanced computer learning algorithms. The first two were based on decision trees, a method where the computer asks a series of yes-or-no questions about the blood values to reach a conclusion. The third was a deep learning model designed specifically for tabular data, which attempts to find complex, non-linear relationships between the numbers. The fourth approach was a "stacking" ensemble, which did not rely on a single algorithm but instead combined the predictions of the other models to make a final decision. The researchers ran these models through a rigorous testing process, ensuring that the data used to teach the computer was never the same data used to test it, a step necessary to prove the system works on new, unseen patients.

The results showed that the combined approach was the most effective. The stacking ensemble model, which pooled the strengths of the other algorithms, achieved an accuracy of 99.89 percent in correctly identifying the three groups. It successfully identified 94.7 percent of the actual disease and carrier cases, a critical measure in medicine where missing a diagnosis can have serious consequences. The other models performed well as well, with the decision-tree-based models coming in close behind, but the combined model consistently outperformed them all. The researchers also used a method to understand exactly how the computer was making its decisions. They found that the system relied most heavily on five specific biomarkers: the level of a hemoglobin type called HbA2, the average size of the red blood cells, the amount of hemoglobin inside those cells, the total count of red blood cells, and the level of fetal hemoglobin.

These findings are significant because they align perfectly with established medical knowledge. Doctors have long known that elevated levels of HbA2 and smaller-than-average red blood cells are the hallmarks of thalassemia carriers, while changes in fetal hemoglobin often signal more severe forms of the disease. The fact that the artificial intelligence independently identified these same factors as the most important gives doctors confidence that the system is not just guessing but is actually learning the biological rules of the disease. The study also highlighted that while deep learning models are powerful, they may require even larger datasets to outperform the more traditional decision-tree methods when working with structured medical data.

The researchers acknowledge that their system is not yet a perfect solution for every situation. The data came from a single region in India, and the genetic patterns of the disease can vary in other parts of the world. Additionally, the system is complex, and the most accurate model is harder to explain than a single decision tree. However, the study demonstrates that it is possible to build a highly accurate, automated screening tool that can handle the messy reality of real-world medical data. By combining different types of blood tests and using a smart combination of computer algorithms, this framework offers a promising path toward faster, more reliable screening for thalassemia, potentially helping to identify carriers and patients earlier and more effectively than current methods allow.