Machine Learning and Large Language Models for Maternal Cardiometabolic Risk Prediction Across Harmonised Public Pregnancy Datasets

Pregnancy is a profound physiological event that acts as a rigorous stress test for a woman's heart and metabolism. During these nine months, the body must manage increased blood volume, shifting hormone levels, and the metabolic demands of a growing fetus. For many, this process proceeds smoothly, but for others, it reveals underlying vulnerabilities. Conditions such as high blood pressure, irregular blood sugar, and low levels of iron in the blood can emerge or worsen, creating a complex web of risks that threaten both the mother and the baby. These complications are not isolated events; they are often interconnected, and their presence can signal a higher likelihood of difficult deliveries, premature birth, or long-term health issues for the mother later in life. Identifying these risks early allows doctors to intervene, monitor more closely, and tailor care to prevent serious outcomes.

For decades, medical professionals have relied on specific thresholds and individual indicators to spot these dangers. A blood pressure reading above a certain number or a glucose level outside a normal range triggers a warning. However, the human body does not always follow simple, linear rules. The interaction between a woman's age, her weight gain, her history of past pregnancies, and her current vital signs creates a pattern that is difficult to capture with a single checklist. This complexity has led researchers to explore whether computers can learn to recognize these subtle, non-linear patterns better than traditional methods. The hope is that by feeding vast amounts of clinical data into advanced algorithms, machines could predict who is most at risk, even before symptoms become obvious.

A recent study set out to test this idea using two different approaches: traditional machine learning and the newer, more conversational large language models. The researchers began by gathering data from two publicly available collections of maternal health records. One dataset contained general information on pregnant women, while the other focused specifically on a condition called preeclampsia, a serious disorder involving high blood pressure. To make these two different sources work together, the team had to translate their variables into a common language. They aligned measurements like blood pressure, hemoglobin levels, and gestational age so that a reading in one dataset meant the same thing in the other. They also standardized the risk labels, grouping women into low, medium, and high-risk categories based on their existing medical records.

The first phase of the research involved training a set of computer models on the first dataset. These models were tasked with looking at a woman's clinical profile and predicting her risk category. The researchers tested several different types of algorithms, including decision trees and support vector machines, to see which one learned the patterns most effectively. One model, known as HistGradientBoosting, emerged as the strongest performer. When tested against the data it had learned from, it correctly identified the risk level for nearly 88 percent of the cases. It demonstrated strong internal discrimination in separating high-risk patients from the rest, with an area under the receiver operating characteristic curve of 0.989 and an area under the precision-recall curve of approximately 0.974. The model also identified which factors mattered most, highlighting that distance to a hospital, blood glucose levels, and weight gain during pregnancy were key drivers in its decisions.

However, the true test of any medical prediction tool is not how well it works on the data it was trained on, but how well it works on new, unseen data. This is where the study took a critical turn. The researchers took their best-performing model and applied it directly to the second dataset, which came from a different source and contained different patients. They did not retrain the model or adjust its settings; they simply asked it to make predictions based on the six variables that both datasets shared. The results were starkly different. The model's accuracy dropped significantly, correctly classifying only about 39 percent of the cases. More troublingly, it failed to identify a single woman who was actually high-risk in this new group. Instead, it labeled all 26 high-risk women as low or medium risk. This failure suggested that the patterns the model had learned were specific to the first group of patients and did not translate to the second, likely because the two groups differed in fundamental ways, such as the average age of the mothers or the timing of their blood tests.

Recognizing that traditional models might struggle with this lack of generalizability, the researchers turned to large language models. These are the same types of artificial intelligence systems that can write essays, answer questions, and hold conversations. The team converted the structured medical records into plain English narratives, describing each patient's history and current condition in a few sentences. They then asked two different language models, DeepSeek and Qwen, to read these stories and predict the risk level. They tried various methods of prompting the models, including giving them examples of how to solve similar problems before asking them to solve a new one. The best-performing configuration, which used a specific type of reasoning prompt, managed to correctly identify the risk level for about 58 percent of the cases. While this was better than the traditional model's performance on the new data, it was still far from perfect. The model showed a clear bias: it successfully identified all four high-risk cases in the small test group but failed to identify any of the low-risk cases, essentially guessing that everyone was at least somewhat at risk.

The researchers also looked at how these different systems explained their decisions. They compared the factors that the traditional machine learning model considered important with the keywords the language models used in their written explanations. Both systems agreed that blood pressure, blood sugar, and hemoglobin levels were critical indicators of risk. However, they diverged on other factors. The traditional model placed significant weight on how far a patient lived from a hospital and how much weight she had gained, likely because these were strong statistical signals in the data. The language models, conversely, rarely mentioned these logistical and physical factors. Instead, they focused more heavily on a patient's past medical history and previous complications, drawing on the clinical knowledge embedded in their training. This difference highlights a gap between what the data statistically suggests and what the models find clinically plausible.

The study concludes that while these tools show promise, they are not yet ready for clinical use. The traditional machine learning model performed well within the environment it was built for but could not adapt to a new population, a common hurdle in medical artificial intelligence. The large language models showed an ability to read case notes and generate human-readable explanations, but their predictions were unstable and prone to bias. The researchers emphasize that the value of this work lies not in producing a single perfect predictor, but in demonstrating the difficulty of moving a model from one dataset to another. They suggest that for these technologies to become useful in real-world hospitals, they must be tested on much larger, more diverse groups of patients and validated against actual health outcomes rather than just existing labels. Until then, these systems remain experimental tools that can organize information and suggest possibilities, but they cannot yet replace the judgment of a clinician.