Learning Under Extreme Label Imbalance in EHRs: A Dependency-Aware Loss for Multi-Label Classification

Iris Szu-Szu Ho, Lars Werne, Konrad Rawlik, Bruce Guthrie, Sohan Seth
Proceedings of the 7th Conference on Health, Inference, and Learning, PMLR 333:880-904, 2026.

Abstract

Extreme multi-label next-visit diagnosis forecasting from electronic health records is dominated by label sparsity. Each visit contains only a handful of positive ICD-10 codes among thousands of candidates, yet codes are strongly correlated through comorbidity structure. In this regime, standard element-wise objectives (such as focal, and class-balanced loss) often maximize sensitivity at the cost of severe precision degradation, producing clinically impractical alert volumes. We propose an architecture-compatible dependency-aware ranking loss that (i) reweights per-code correctness under severe imbalance, (ii) aggregates errors with rank-based emphasis on the hardest labels, and (iii) regularizes predictions with a learned pairwise dependency term in the output space. Using an EHR Transformer backbone, we evaluate on the CPRD cohort ($V{=}1{,}538$ codes), benchmarking loss functions on 200{,}000 patients and validating scalability up to 3.2 million. The proposed objective shifts the precision–recall trade-off toward fewer false positives while maintaining competitive sensitivity, and preserves overall ranking quality (PRC–AUC comparable to weighted BCE). In addition, it yields an auditable population-level dependency matrix summarizing learned co-occurrence structure. These results suggest that explicit output-space structure can improve the precision–recall trade-off in sparse, high-dimensional next-visit diagnosis prediction from EHRs.

Cite this Paper

BibTeX
@InProceedings{pmlr-v333-ho26a, title = {Learning Under Extreme Label Imbalance in EHRs: A Dependency-Aware Loss for Multi-Label Classification}, author = {Ho, Iris Szu-Szu and Werne, Lars and Rawlik, Konrad and Guthrie, Bruce and Seth, Sohan}, booktitle = {Proceedings of the 7th Conference on Health, Inference, and Learning}, pages = {880--904}, year = {2026}, editor = {Healey, Elizabeth and Fries, Jason and Pollard, Tom and Tang, Shengpu and Zink, Anna and Hartvigsen, Tom and Agrawal, Monica and Finlayson, Sam and Glicksberg, Benjamin and Beaulieu-Jones, Brett and Wang, Kai and Fontalvo, Daseyra and Sarker, Tasmie and Chen, Irene and Alsentzer, Emily}, volume = {333}, series = {Proceedings of Machine Learning Research}, month = {29--30 Jun}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v333/main/assets/ho26a/ho26a.pdf}, url = {https://proceedings.mlr.press/v333/ho26a.html}, abstract = {Extreme multi-label next-visit diagnosis forecasting from electronic health records is dominated by label sparsity. Each visit contains only a handful of positive ICD-10 codes among thousands of candidates, yet codes are strongly correlated through comorbidity structure. In this regime, standard element-wise objectives (such as focal, and class-balanced loss) often maximize sensitivity at the cost of severe precision degradation, producing clinically impractical alert volumes. We propose an architecture-compatible dependency-aware ranking loss that (i) reweights per-code correctness under severe imbalance, (ii) aggregates errors with rank-based emphasis on the hardest labels, and (iii) regularizes predictions with a learned pairwise dependency term in the output space. Using an EHR Transformer backbone, we evaluate on the CPRD cohort ($V{=}1{,}538$ codes), benchmarking loss functions on 200{,}000 patients and validating scalability up to 3.2 million. The proposed objective shifts the precision–recall trade-off toward fewer false positives while maintaining competitive sensitivity, and preserves overall ranking quality (PRC–AUC comparable to weighted BCE). In addition, it yields an auditable population-level dependency matrix summarizing learned co-occurrence structure. These results suggest that explicit output-space structure can improve the precision–recall trade-off in sparse, high-dimensional next-visit diagnosis prediction from EHRs.} }
Endnote
%0 Conference Paper %T Learning Under Extreme Label Imbalance in EHRs: A Dependency-Aware Loss for Multi-Label Classification %A Iris Szu-Szu Ho %A Lars Werne %A Konrad Rawlik %A Bruce Guthrie %A Sohan Seth %B Proceedings of the 7th Conference on Health, Inference, and Learning %C Proceedings of Machine Learning Research %D 2026 %E Elizabeth Healey %E Jason Fries %E Tom Pollard %E Shengpu Tang %E Anna Zink %E Tom Hartvigsen %E Monica Agrawal %E Sam Finlayson %E Benjamin Glicksberg %E Brett Beaulieu-Jones %E Kai Wang %E Daseyra Fontalvo %E Tasmie Sarker %E Irene Chen %E Emily Alsentzer %F pmlr-v333-ho26a %I PMLR %P 880--904 %U https://proceedings.mlr.press/v333/ho26a.html %V 333 %X Extreme multi-label next-visit diagnosis forecasting from electronic health records is dominated by label sparsity. Each visit contains only a handful of positive ICD-10 codes among thousands of candidates, yet codes are strongly correlated through comorbidity structure. In this regime, standard element-wise objectives (such as focal, and class-balanced loss) often maximize sensitivity at the cost of severe precision degradation, producing clinically impractical alert volumes. We propose an architecture-compatible dependency-aware ranking loss that (i) reweights per-code correctness under severe imbalance, (ii) aggregates errors with rank-based emphasis on the hardest labels, and (iii) regularizes predictions with a learned pairwise dependency term in the output space. Using an EHR Transformer backbone, we evaluate on the CPRD cohort ($V{=}1{,}538$ codes), benchmarking loss functions on 200{,}000 patients and validating scalability up to 3.2 million. The proposed objective shifts the precision–recall trade-off toward fewer false positives while maintaining competitive sensitivity, and preserves overall ranking quality (PRC–AUC comparable to weighted BCE). In addition, it yields an auditable population-level dependency matrix summarizing learned co-occurrence structure. These results suggest that explicit output-space structure can improve the precision–recall trade-off in sparse, high-dimensional next-visit diagnosis prediction from EHRs.
APA
Ho, I.S., Werne, L., Rawlik, K., Guthrie, B. & Seth, S.. (2026). Learning Under Extreme Label Imbalance in EHRs: A Dependency-Aware Loss for Multi-Label Classification. Proceedings of the 7th Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 333:880-904 Available from https://proceedings.mlr.press/v333/ho26a.html.

Related Material