Tridirectional Transfer Learning for Predicting Gastric Cancer Morbidity.

Area of Science:

• Environmental Health
• Epidemiology
• Machine Learning

Background:

• A prior deep learning model predicted gastrointestinal infection morbidity using environmental pollutants in central China.
• Adapting predictive models for new diseases or regions is crucial for public health preparedness.

Purpose of the Study:

• To adapt a deep learning model for predicting disease morbidity across different diseases and geographical regions.
• To develop a tridirectional transfer learning approach for enhanced disease prediction capabilities.

Main Methods:

• A combined univariate regression and multivariate Gaussian model was developed to link source and target disease morbidity with pollutant features.
• Mapping-based deep transfer learning extended the model to predict source disease morbidity in new regions.
• Source region model patterns were applied to derive new models for target disease prediction in target regions, using gastric cancer as the target disease.

Main Results:

• The transfer learning approach achieved over 80% prediction accuracy in the source region.
• Prediction accuracy reached up to 78% in the target regions, despite limited labeled samples.
• The model effectively predicted gastric cancer morbidity in both source and multiple target regions.

Conclusions:

• The proposed tridirectional transfer learning method successfully adapts disease prediction models to new contexts (different diseases/regions).
• This approach demonstrates high accuracy even with sparse data, offering a valuable tool for medical preparedness.
• The findings highlight the potential of transfer learning in environmental health and epidemiological forecasting.