Do We Really Want Medical AI to Be Correct?

What if the most accurate medical AI model wasn't necessarily the one you'd want making the decision?

That question became the unexpected result of my experiment.

When doctors need to determine whether an unusual area in the breast could be cancerous, they may perform a biopsy: a procedure that removes a small sample of tissue for examination. Specialists can then study the cells and look for patterns that distinguish cancerous tissue from noncancerous tissue.

Could a computer learn to recognize some of those patterns? And if it makes mistakes, should every mistake count the same?

Machine learning is a way of teaching computers to recognize patterns in data. In this project, I used four machine-learning models to analyze data from 569 digitized breast-tumor samples. Each sample contained measurements of cell nuclei, including their size, shape, texture, and irregularity. The models used these measurements to classify each sample as malignant, meaning cancerous, or benign, meaning noncancerous.

I began with a straightforward goal: to find the model that classified the samples most accurately. I compared four different models, logistic regression, random forest, support vector machine, and K-nearest neighbors, using several measures of performance.

Logistic regression achieved the highest overall accuracy, correctly classifying 98.83% of the test samples. But K-nearest neighbors identified every malignant sample in the test set, even though its overall accuracy was lower. It also incorrectly flagged more benign samples as malignant.

I had expected the results to reveal one clear winner. Instead, they raised a question I had not begun with: are all incorrect predictions equal?

A false negative happens when a malignant sample is incorrectly labeled benign. A false positive happens when a benign sample is incorrectly flagged as malignant. Both are incorrect predictions, but they can have very different consequences.

Imagine two hypothetical medical AI tools. One is less likely to miss a cancerous sample, but it could also produce more false alarms. The other is more accurate overall, but it could miss a malignant sample. Which risk should matter more: missing a malignant sample, or creating more false alarms?

My experiment measured how often these mistakes occurred in the dataset, but it didn't measure their emotional, financial, or medical effects on real patients. Still, the different results from the models made the trade-off impossible to ignore. In medical AI, the important question may not be simply "how often is the model correct?" but also "what happens when it is wrong?"

The data also showed patterns that helped distinguish cancerous samples from noncancerous ones. In the Random Forest analysis, measurements related to the size and irregularity of cell nuclei ranked among the strongest predictors of malignancy. These measurements resemble features that specialists examine when studying cells under a microscope. This suggests the models are picking up on real, clinically relevant signals, but it doesn't tell us which model should be trusted in patient care.

The importance of different errors also depends on what a medical AI tool is designed to do. For example, a tool built to reduce missed malignant cases might place more weight on finding those cases, even if it produces more false alarms. A tool designed for a different purpose might place greater emphasis on reducing unnecessary alerts. The appropriate balance depends on the consequences of each type of error.

This question matters beyond this dataset. As AI becomes more involved in decisions that affect people's health, young people may encounter these systems as patients, family members, or future healthcare professionals. Knowing that a model can be "98% accurate" is not enough. We also need to ask what happens in the remaining 2%, and who bears the consequences when the system is wrong.

My study was a proof of concept based on a public dataset. Before similar models could support real patient care, they would need to be tested on larger and more diverse datasets and evaluated in real healthcare settings. The results do not show that K-nearest neighbors is universally better than logistic regression, or that either model is ready to diagnose patients.

The main lesson from my project is that choosing a medical AI tool requires looking beyond a single measure of accuracy. Researchers and healthcare professionals must also ask what kinds of mistakes a model makes, how serious those mistakes could be, and what the tool is designed to do.

AI can identify patterns and measure trade-offs. But deciding which mistakes matter most remains a human responsibility.

Original research: Keerat Singh — The Role of Artificial Intelligence and Bioinformatics in Accelerating Early Disease Diagnosis and Patient Care