FlatASCEND: Autoregressive Model for Clinical Sequence Generation
The FlatASCEND model, comprising 14.5 million parameters, is an autoregressive system that produces multi-step clinical trajectories tailored to patient conditions in response to intervention tokens. It employs flat composite tokens alongside a zero-inflated log-normal time head. Although standard distributional metrics (Jaccard 0.889-0.954) do not differentiate it from basic baselines, its significance lies in generating conditions based on patient-specific prefixes. An ablation study on prompt-shuffling indicates that patient-specific conditioning enhances mechanistic pharmacological effects (2.0-2.2x for steroid to glucose and diuretic to potassium), while confounding-driven associations remain stable (0.9x for insulin to glucose). A framework for incident users evaluates directional consistency with existing pharmacological knowledge.
Key facts
- FlatASCEND is a 14.5M-parameter autoregressive clinical sequence model.
- It uses flat composite tokens and a zero-inflated log-normal time head.
- Standard distributional metrics (Jaccard 0.889-0.954) do not distinguish FlatASCEND from trivial baselines.
- Model's value lies in conditional generation from patient-specific prefixes.
- Prompt-shuffle ablation shows patient-specific conditioning amplifies mechanistic pharmacological effects (2.0-2.2x for steroid to glucose, diuretic to potassium).
- Confounding-driven associations remain unchanged (0.9x for insulin to glucose).
- An incident-user framework assesses directional consistency against prior pharmacological knowledge.
- The model generates patient-conditioned multi-step trajectories that respond to intervention tokens.
Entities
—