Learning Deep Temporal Representations of Socioeconomic Deprivation

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Public Health

Background:

• Socioeconomic status significantly impacts health outcomes and disease progression.
• Traditional indices of socioeconomic deprivation may not fully capture complex societal disparities.
• Accurate quantification of socioeconomic factors is crucial for personalized medicine and public health interventions.

Purpose of the Study:

• To evaluate if AI-derived embeddings from a recurrent neural network (RNN) can more effectively capture socioeconomic deprivation than established expert-based indices.
• To compare the predictive performance of RNN embeddings against the German Index of Socioeconomic Deprivation (GISD) for cancer patient survival.

Main Methods:

• Development of RNN embeddings trained on features of a deprivation index.
• Utilizing the German Index of Socioeconomic Deprivation (GISD) as a benchmark for comparison.
• Conducting survival analyses on three independent cancer cohorts (thyroid, breast, and lung cancer) to assess predictive performance.

Main Results:

• RNN embeddings consistently demonstrated superior predictive performance compared to the GISD score.
• Significant improvements in overall survival prediction were observed across all evaluated cancer types.
• The most substantial enhancement was noted in thyroid carcinoma (c-index +6.85%), followed by breast cancer (4.46%) and lung cancer (2.19%).

Conclusions:

• AI-based embeddings, particularly from RNNs, offer a more nuanced and powerful quantification of socioeconomic deprivation.
• This advanced AI methodology enhances the prediction of cancer patient survival, outperforming traditional socioeconomic indices.
• The findings suggest a promising role for AI in refining health disparity assessments and improving prognostic models in oncology.