How Can AI Improve Fetal Health Diagnosis?

During pregnancy, monitoring the baby's well‑being is crucial. A key tool is fetal cardiotocography (CTG), which records fetal heartbeats and uterine contractions. However, CTG data are often imbalanced: rare conditions appear far less often, making it hard for traditional models to recognize them correctly. The new study proposes a two‑step strategy. First, an algorithm called MAAL (Adaptive Adjustment of Misclassification Loss) dynamically adjusts loss‑function weights, focusing more on samples the model misclassifies. In practice, if the model confuses a class, that error’s weight increases, pushing the network to pay more attention. Second, a primary‑secondary correction network (MAC‑NET) is introduced. After the primary model makes its predictions, the secondary network reviews the wrong cases to correct them. Tests on a public fetal health dataset (UCI) showed an accuracy of 99.39 %, far surpassing previous methods. Moreover, the same architecture performed well on other imbalanced datasets, demonstrating its versatility. These results suggest that, thanks to adaptive learning techniques, AI can lower the risk of misdiagnosis and help doctors manage high‑risk pregnancies early.

The MAAL model relies on a dynamically weighted loss function. For each training batch, the error rate per class is computed; classes with higher error receive a larger scaling factor. This approach resembles focal loss but features a more reactive update mechanism, as weights adjust each epoch based on current performance. MAC‑NET, on the other hand, is built as two sequential modules: the first is a standard CNN extracting temporal features from the CTG series, while the second is a lighter network (often an MLP) trained solely on samples misclassified by the first module. Correction occurs via residual learning, where the second module's output is added to the original output, refining the final decision. Statistically, this reduces bias toward majority classes, improving ROC curves and AUC on imbalanced sets. Moreover, the MAAL‑MAC‑NET combination maintains good generalization, avoiding the over‑fitting typical of simple oversampling methods. Implementation-wise, training takes about 2‑3 times longer than a single network, but the accuracy gain (over 5 percentage points) justifies the extra computational cost, especially in clinical settings where precision is critical.