Abstract
The efficacy of machine learning in clinical diagnostics is frequently compromised by "label noise"—incorrectly categorized medical data resulting from human error or subjective bias. This paper introduces a robust disease prediction framework utilizing an Adaptive Modular Neural Network (AMNN) coupled with a dynamic truncated loss mechanism. By employing a "divide and conquer" strategy through density peak clustering and reducing the influence of high-loss samples during training, the proposed model achieves a stable 97% accuracy even when training data contains up to 40% noise. Experimental validation on a stroke screening dataset demonstrates that this approach significantly outperforms standard deep learning and ensemble models in maintaining predictive integrity under adverse data conditions.
Introduction
Machine learning models are fundamentally dependent on the quality of their training data. In the medical domain, "label noise" presents a substantial challenge. Clinical records are often manually labeled by practitioners with varying levels of experience, leading to subjective biases. Furthermore, diagnostic complexities and coding errors can result in datasets where high-risk patients are mislabeled as low-risk, or vice versa. Standard deep learning models tend to overfit to this noise, learning erroneous patterns that degrade their real-world predictive power. This research addresses the need for AI tools that can distinguish between reliable clinical data and suspected noise to ensure diagnostic accuracy in practical settings.
Methodology
The proposed framework utilizes a two-pronged architectural approach to mitigate the effects of label noise:
1. Adaptive Modular Neural Network (AMNN)
Rather than relying on a single monolithic network, the AMNN adopts a modular strategy. It utilizes a density peak clustering algorithm to partition the training data into distinct clusters based on patient characteristics. Specific "sub-networks" are then assigned to learn from these clusters. This modularity allows the system to adapt to the unique features of different patient subgroups, making it easier to identify outliers that do not fit the density profile of a specific cluster.
2. Dynamic Truncated Loss and Fuzzy Membership
To prevent the model from being "led astray" by incorrect labels, a dynamic truncated loss function is implemented. This function identifies "suspected" noisy samples—those exhibiting unusually high loss during the early stages of training—and reduces their gradient contribution. This approach leverages the observation that neural networks typically learn "pure" patterns before overfitting to noise. Additionally, fuzzy set theory is applied to determine the membership degree of data points to specific cluster centers, refining the task decomposition process.
Experimental Setup
The model was evaluated using a comprehensive stroke screening dataset comprising 898,171 records. The data was categorized into three risk levels: low-risk, moderately dangerous, and high-risk.
To test the robustness of the proposed method, researchers manually injected label noise into the dataset at ratios ranging from 5% to 40%.
Results and Discussion
On perfectly labeled "clean" data, the proposed model and standard deep neural networks (DNNs) both achieved high performance, with accuracy and F1 scores exceeding 0.98. However, the true strength of the AMNN framework became apparent as the level of label noise increased.
As shown in the comparative analysis, standard DNN accuracy dropped to 80.90% at 40% noise, while popular ensemble methods like LightGBM and XGBoost fell to approximately 78%. In contrast, the proposed robust method (referred to as the "Textual algorithm" in experimental logs) maintained a consistent accuracy of 97% across all noise levels. This stability indicates that the modular approach and truncated loss effectively filter out corrupted data, prioritizing robustness over raw fitting power.
Conclusion
This research demonstrates that an adaptive, modular approach to neural network design can significantly enhance the reliability of disease prediction models in the presence of label noise. While standard models suffer significant performance degradation as data quality drops, the proposed AMNN framework provides a stable and accurate diagnostic tool. Future work will focus on validating this method across a broader range of diseases and optimizing the computational complexity associated with the modular sub-network selection process.
Summary updated