Representation Effects in Child and Youth Mental Health Emergency Readmission Predictions

Cassandra Czobit, Hirad Daneshvar, Reza Samavi, Thomas Doyle, Laura Duncan, Paulo Pires, Roberto Sassi
Proceedings of the The 39th Canadian Conference on Artificial Intelligence, PMLR 318:1028-1035, 2026.

Abstract

Predicting mental health–related emergency department readmission in youth remains challenging, and the role of data representation is underexplored. Using the National Survey on Drug Use and Health (ages 12–18), we compare three representations: (1) structured tabular features, (2) template-generated clinical text, and (3) LLM-derived sentence embeddings. Classical models are trained on tabular data and embeddings, while LLMs are applied to text. Results show that tabular features consistently yield the best and most stable performance. Templated text introduces a representational bottleneck and is less robust under distribution shift, while embeddings preserve some semantics but do not outperform tabular inputs. Representation choice is thus critical for predictive performance.

Cite this Paper

BibTeX
@InProceedings{pmlr-v318-czobit26a, title = {Representation Effects in Child and Youth Mental Health Emergency Readmission Predictions}, author = {Czobit, Cassandra and Daneshvar, Hirad and Samavi, Reza and Doyle, Thomas and Duncan, Laura and Pires, Paulo and Sassi, Roberto}, booktitle = {Proceedings of the The 39th Canadian Conference on Artificial Intelligence}, pages = {1028--1035}, year = {2026}, editor = {Bouzar-Benlabiod, Lydia and Leung, Carson}, volume = {318}, series = {Proceedings of Machine Learning Research}, month = {25--29 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v318/main/assets/czobit26a/czobit26a.pdf}, url = {https://proceedings.mlr.press/v318/czobit26a.html}, abstract = {Predicting mental health–related emergency department readmission in youth remains challenging, and the role of data representation is underexplored. Using the National Survey on Drug Use and Health (ages 12–18), we compare three representations: (1) structured tabular features, (2) template-generated clinical text, and (3) LLM-derived sentence embeddings. Classical models are trained on tabular data and embeddings, while LLMs are applied to text. Results show that tabular features consistently yield the best and most stable performance. Templated text introduces a representational bottleneck and is less robust under distribution shift, while embeddings preserve some semantics but do not outperform tabular inputs. Representation choice is thus critical for predictive performance.} }
Endnote
%0 Conference Paper %T Representation Effects in Child and Youth Mental Health Emergency Readmission Predictions %A Cassandra Czobit %A Hirad Daneshvar %A Reza Samavi %A Thomas Doyle %A Laura Duncan %A Paulo Pires %A Roberto Sassi %B Proceedings of the The 39th Canadian Conference on Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Lydia Bouzar-Benlabiod %E Carson Leung %F pmlr-v318-czobit26a %I PMLR %P 1028--1035 %U https://proceedings.mlr.press/v318/czobit26a.html %V 318 %X Predicting mental health–related emergency department readmission in youth remains challenging, and the role of data representation is underexplored. Using the National Survey on Drug Use and Health (ages 12–18), we compare three representations: (1) structured tabular features, (2) template-generated clinical text, and (3) LLM-derived sentence embeddings. Classical models are trained on tabular data and embeddings, while LLMs are applied to text. Results show that tabular features consistently yield the best and most stable performance. Templated text introduces a representational bottleneck and is less robust under distribution shift, while embeddings preserve some semantics but do not outperform tabular inputs. Representation choice is thus critical for predictive performance.
APA
Czobit, C., Daneshvar, H., Samavi, R., Doyle, T., Duncan, L., Pires, P. & Sassi, R.. (2026). Representation Effects in Child and Youth Mental Health Emergency Readmission Predictions. Proceedings of the The 39th Canadian Conference on Artificial Intelligence, in Proceedings of Machine Learning Research 318:1028-1035 Available from https://proceedings.mlr.press/v318/czobit26a.html.

Related Material