Research map / Image generation / Autoregressive and latent generation
ConSequence: Synthesizing Logically Constrained Sequences for Electronic Health Record Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ConSequence, a rule-neuron module with temporal aggregation implemented as matrix multiplications that enforces hard and soft temporal logical constraints in sequential EHR generators (applied to HALO), achieving zero rule violations with improved perplexity and under 13% generation slowdown.
Problem
Synthetic electronic health record generation from sequential generative models often violates domain logical constraints, especially temporal constraints spanning multiple visits, and existing constraint methods are either not guaranteed, unable to handle temporal rules, or computationally slow.
Approach
Encodes entailment rules in conjunctive implicative form, aggregates past visits with a temporal mask, and evaluates antecedents with a threshold rule neuron whose weighted sum is computed as batched matrix multiplication, so the neuron overrides the base model's output (Bernoulli-sampled for soft rules) to guarantee constraint satisfaction.
Key results
- ConSequence yields zero static and temporal rule violations on both datasets (100% valid), matching post-processing, while vanilla HALO has 87.2% and 65.4% valid records.
- ConSequence reduces test perplexity from 24.019 to 23.924 (outpatient) and 13.443 to 12.712 (inpatient).
- ConSequence incurs less than a 13% slowdown in generation speed relative to unconstrained HALO on both datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ConSequence logical constraint enforcement | HALO · Outpatient EHR | % valid: 87.2% (0.002) -> 100% | Helped |
| Add ConSequence logical constraint enforcement | HALO · Inpatient EHR | % valid: 65.4% (0.002) -> 100% | Helped |
| Add ConSequence training with rule-based constraints | HALO · Outpatient EHR test set | perplexity: 24.019 -> 23.924 | Helped |
| Add ConSequence training with rule-based constraints | HALO · Inpatient EHR test set | perplexity: 13.443 -> 12.712 | Helped |
| Use Semantic Loss constraint regularization | HALO · Outpatient EHR test set | perplexity: 24.019 -> 23.995 | Helped |
| Use CCN constraint enforcement | HALO · Inpatient EHR test set | perplexity: 13.443 -> 12.985 | Helped |
| Use MultiPlexNet constraint enforcement | HALO · Outpatient EHR test set | perplexity: 24.019 -> 119.560 | Hurt |
| Use SPL constraint enforcement | HALO · Inpatient EHR test set | perplexity: 13.443 -> 22.794 | Hurt |
Surprising findings
- MultiPlexNet and SPL substantially worsen perplexity relative to vanilla HALO (e.g., MultiPlexNet 119.56 vs 24.02 on outpatient).
- Post-processing achieves full validity but is slower due to repeated generation and rule checking.
Limitations
- Rules were generated data-driven from co-occurrence statistics rather than expert-provided, so clinical validity of rules is not established.
- Some baselines (SPL) were modified for partial compatibility and SPL on outpatient was excluded from speed runs due to time.
- Assumes acyclic rules; cyclic rules are not discussed.
Experimental setup
Datasets: US claims outpatient EHR dataset MIMIC-III
Benchmarks: Outpatient EHR dataset MIMIC-III
Metrics: rule violation count % valid records perplexity training speed generation speed
Baselines: Vanilla HALO Post Processing Semantic Loss CCN MultiPlexNet Semantic Probability Layer
Builds on: HALO Semantic Loss CCN MultiPlexNet Semantic Probability Layer
Closest papers by meaning
- Collaborative Synthesis of Patient Records through Multi-Visit Health State Inference (AAAI 2024 · similarity 0.85)
- How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data (ICLR 2024 · similarity 0.84)
- Improving the Generation and Evaluation of Synthetic Data for Downstream Medical Causal Inference (NeurIPS 2025 · similarity 0.83)
- Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.82)
- Constrained Discrete Diffusion (NeurIPS 2025 · similarity 0.82)
- ORACLE: Optimizing Reasoning Abilities of Large Language Models via Constraint-Led Synthetic Data Elicitation (AAAI 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.