On Entity Embeddings for Ordinal Features as Representation Learning in Recurrence Prediction of Urothelial Bladder...

Area of Science:

• Oncology
• Data Science
• Bioinformatics

Background:

• Urothelial Bladder Cancer (UBC) recurrence is influenced by various factors, posing a significant clinical challenge.
• Machine Learning (ML) shows promise in predicting UBC recurrence, often outperforming traditional methods.
• ML algorithms require numerical input, necessitating effective encoding of categorical data.

Purpose of the Study:

• To evaluate the impact of different encoding strategies on ML prediction quality for UBC recurrence.
• To compare one-hot, ordinal, and entity embedding for encoding ordinal features.

Main Methods:

• An artificial neural network was employed to predict 2-year UBC recurrence.
• Three encoding strategies (one-hot, ordinal, entity embedding) were compared.
• Data from the Cancer Registry Rhineland-Palatinate was utilized.

Main Results:

• Entity embedding achieved superior prediction quality with 84.6% precision and 73.8% accuracy.
• The model demonstrated 68.9% AUC on testing data.
• Entity embedding outperformed one-hot and ordinal encoding.

Conclusions:

• Entity embedding encoding provides a more accurate numerical representation of ordinal features.
• This enhanced representation leads to improved generalizability and prediction quality for UBC recurrence.
• Entity embedding is recommended for ML models dealing with ordinal categorical data in oncology.