Answers are grounded in this paper and cite their sources.
Attention-based sequence models are increasingly proposed for healthcare fraud detection but are rarely compared with strong tabular baselines. We test whether an attention-enhanced CNN-RNN helps on Medicare provider data and whether its attention weights indicate feature importance. We compared nine models, from an LR-RF hybrid and gradient-boosted trees to an MLP, FT-Transformer, CNN-RNN and attention CNN-RNN, over repeated random splits with natural-distribution test sets and validation-tuned thresholds. Dataset A has 100,000 Medicare provider records labelled by a statistical rule on submitted charges; we evaluated it with and without the label-defining charge features. Dataset B has 5,410 providers with fraud labels supplied with the data. We also retrained the attention model with shuffled feature order. On dataset A, gradient-boosted trees achieved a fraud F1 of 0.952. This is much better than the best neural model (0.769) and the attention CNN-RNN (0.750). When charge features were removed, every model’s performance dropped sharply, with scores falling between 0.42 and 0.52. On dataset B, all nine models scored between 0.588 and 0.629. This variation is similar to what is seen between train-test splits. Attention never reliably improved the CNN-RNN, and its weights followed input position rather than features. The large advantage of tree ensembles on rule-labelled data reflects reconstruction of the labelling rule; with labels supplied with the data, architecture choice made little difference. Sequence-based models and attention did not provide reliable improvements on either dataset. Also, attention weights should not be considered as explanations for tabular fraud models.
No takes yet. Share an insight, caveat, or question.
Ullah et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: