Foreseeing Survival through Fuzzy Intelligence: Cognitively-Inspired Incremental Learning model for Breast Cancer Prognosis by Multiomics Data Fusion

High-precision breast cancer prognosis is crucial for early disease identification, avoiding hazardous side-effects of unnecessary therapies, and decreasing mortality rates through personalized and tailored treatment regimens. However, designing a prognosis model continues to be challenging, given the intricate relationship between distinct genetic attributes, varied clinical results of drug therapies, the noisy nature of gene expressions, and the high-class imbalance seen in multimodal cancer data. Furthermore, because labeled omics data collection is costly and requires highly-trained experts, the data available is very limited. This makes the design of the conventional machine and deep learning models incredibly challenging as they require large quantities of data for learning the underlying intricate patterns and would otherwise overfit, decreasing model precision. Moreover, all present models suffer from a ‘closed world assumption.’ These models, once trained, cannot be updated in real-time (when more omics data is available in the future) without a complete re-training. The present study is the first to introduce the ‘Fuzzy’ way towards Breast cancer prognosis, framing the task as an incremental learning problem. The proposed approach allows the model to continually update its learned feature space on a non-stationary multimodal data stream emulating the human brain’s remarkable quality to learn over time. We demonstrate the model’s ability to learn complex relationships between different multimodal attributes, training on severely imbalanced and limited data by mapping it to a high-dimensional ‘fused’ feature space. The proposed model surpasses state-of-the-art machine learning (ML) models significantly. These results suggest that prediction through ‘fuzzy intelligence’ is a promising approach towards breast cancer prognosis.

Illustration of the (a) Conventional prognosis models with closed world assumption, (b) The required re-training for conventional leaned feature space update (c) Proposed real-time update of learned fused feature space.

Accurate breast cancer prognosis is essential for identifying the disease early, avoiding unnecessary and harmful treatments, and reducing mortality rates with personalized treatment plans. However, designing a prognosis model continues to be challenging. However, creating a reliable prognosis model is challenging due to the complex interplay of various genetic factors, diverse outcomes of drug therapies, the unpredictable nature of gene expressions, and the imbalanced distribution of data in multimodal cancer datasets.

Even machine learning (ML) models face difficulties because collecting labeled omics data is expensive and requires specialized expertise, limiting the available data. This scarcity makes traditional ML and deep learning models difficult to design, as they typically rely on large datasets to understand intricate patterns and can overfit without sufficient data, leading to reduced accuracy. Additionally, existing models operate under a closed-world assumption, meaning they cannot adapt to new data in real-time without undergoing retraining.

This study introduces a novel 'Fuzzy' approach to breast cancer prognosis, treating the problem as an incremental learning challenge. This method enables the model to continuously update its learned features using a dynamic multimodal data stream, mimicking the human brain's ability to learn over time. Our model demonstrates its capability to understand complex relationships among different multimodal attributes, even when trained on limited and imbalanced data, by mapping it to a high-dimensional 'fused' feature space.

In conclusion, our model can help to increase prognosis accuracy by 16.81% on just training on 300 samples without losing consistency when training dataset reduces/ increases. Our model’s ability to adapt to distribution shifts and its robust performance in limited data scenarios position it as a versatile and enduring tool. The present study further suggests that using fuzzy intelligence holds promise for improving breast cancer prognosis.

Effect of variation in hyperbox (H) expansion coefficient on (a) Number of hyperbox (H) formed during model training (b) Model training time (sec).