EviGen: Predictive Evidence Scaffolding for Verifiable Clinical Rationale Generation

AI summary

Overview

Research area: Clinical natural language processing and retrieval-augmented generation over longitudinal electronic health records (EHRs).

Technical level: Intermediate. The framework combines retrieval architectures, attribution methods, and LLM prompting, but the core ideas are explainable without deep background in any single one.

Scope: The paper introduces EviGen, a three-layer pipeline that selects clinically predictive evidence from multi-year patient records, generates citation-grounded clinical rationales, and verifies each reasoning step—demonstrated on mortality, autism, and ADHD prediction tasks.

What This Paper Is About

Patient records span years of notes, codes, labs, and procedures, and the evidence needed to predict a clinical outcome is often sparse, indirect, and spread across time. Feeding an entire record to an LLM is expensive and unreliable—models miss buried evidence and fabricate claims, and clinicians cannot practically audit every generated statement against a 100,000-token history. EviGen's goal is to make clinical AI reasoning both accurate and traceable: it retrieves only the observations that actually predict the outcome, forces the LLM to quote and cite each one, and then flags unreliable reasoning steps for human review.

Key Contributions

1. A three-layer framework for verifiable clinical rationale generation over full longitudinal EHR histories—to the authors' knowledge the first system designed for this setting rather than for single-note or discharge-summary inputs.

2. Learnable queries as evidence detectors. Instead of matching text to a query, EviGen trains query vectors against outcome labels so they retrieve evidence by predictive utility, and gates them per patient so only the queries relevant to that patient's profile activate—enabling fixed-cost retrieval over arbitrarily long records.

3. Evidence-scaffolded generation. The rationale format requires each reasoning step to quote a supporting passage verbatim, cite its source record ID, and carry a signed attribution score, making the entire rationale auditable without reading the full EHR.

4. A process-supervised verifier that scores generated rationales step by step, assigns type-aware error flags (e.g., hallucination, factual inaccuracy), and routes suspicious steps to clinician review.

5. Empirical validation across three tasks, showing gains in prediction performance, rationale faithfulness, and clinical reviewer preference over full-context and RAG baselines.

Main Findings

• Prediction performance improves across the board. EviGen beats every baseline on all four reported metrics. The largest gaps are +10.4 points in mortality accuracy (0.800 vs. 0.696 for the best baseline) and +8.9 AUC points on ADHD (0.704 vs. 0.615). Mortality AUC reaches 0.871.

• Baselines are poorly calibrated even when they discriminate. Llama-3.1-8B/70B and GPT-4o-mini all post mortality accuracies below 0.5 despite AUCs of 0.69–0.80. EviGen's Expected Calibration Error is 0.042, compared with 0.200–0.536 for zero-shot and RAG baselines.

• RAG does not reliably beat full-context prompting. The paper finds no consistent advantage for standard query-based retrieval, which the authors attribute to its reliance on textual relevance rather than predictive value.

• Faithfulness gains are large and consistent. EviGen achieves the highest joint source-grounding-plus-reasoning-consistency pass rate on 11 of 12 backbone–dataset cells, with margins of 0.6 to 45.1 points. For MIMIC-IV mortality with Llama-3.1-8B, the joint pass rate rises from 6.1% (full context) and 13.7% (RAG) to 58.8%—actually exceeding Llama-3.1-70B with EviGen (56.3%), meaning the evidence scaffold lets a small model match a large one.

• Verbatim quoting is the bottleneck for baselines. Source grounding (Stage 1) is much harder for baselines than reasoning consistency (Stage 2), partly because EHR notes use heavy abbreviations and specialized terms that models tend not to reproduce exactly. Stage 2 pass rates are inflated for baselines by survivor bias—only easily quotable passages reach that check.

• Reasoning models trade quote fidelity for reasoning quality. Qwen3-32B shows low Stage 1 but near-perfect Stage 2 pass rates, consistent with findings that chain-of-thought can degrade instruction following. EviGen sidesteps this trade-off by supplying reasoning support through the scaffold.

• Both design choices in generation matter independently. A matched 2×2 experiment shows the evidence scaffold adds 9.7–21.2 points to the joint pass rate, and switching from similarity-ranked to attribution-ranked evidence adds a further 3.4–11.6 points.

• Attribution ranking surfaces genuinely predictive evidence. When forced to predict from five chunks ranked by Integrated Gradients versus five ranked by query similarity, EviGen's predictor scores 0.868 vs. 0.661 AUC, and the zero-shot LLM backbones gain 13.8–21.4 points—indicating the top-ranked chunks carry transferable predictive signal, not predictor-specific artifacts.

• Supervised baselines confirm the gains are not just from label access. EviGen outperforms fine-tuned QLoRA variants and IRIS (the same architecture with fixed, non-patient-conditioned queries) on all three datasets, with the gap widening to 2.7–3.5 points on the longer, sparser real-EHR autism and ADHD data.

• Compute cost is dramatically lower. EviGen trains in under 30 minutes on a single A5000 Ada GPU, versus 35–80 GPU-hours for 32K-context QLoRA on H200 GPUs, and its cost does not scale with input length.

• Clinical reviewers preferred EviGen. In a pilot with 7 medical students, EviGen rationales scored 3.49/5 on average versus 3.34 for a standard LLM rationale and 2.71 for the prediction with evidence pack alone. Six of seven reviewers chose EviGen as their preferred method. EviGen led on informativeness, actionability, and trust but trailed the simpler LLM rationale on usability—likely because attribution scores and verifier output add interpretive burden.

Methodology in Plain English

The system works in three sequential stages, with observation IDs preserved across all three so the final output remains traceable to the source record.

Layer 1 — Evidence Selection. The patient's notes and ICD codes are embedded with a shared text encoder. A set of learnable query vectors—trained end-to-end using clinical outcome labels extracted from structured EHR fields, so no human annotation is needed—each specialize in detecting a distinct type of predictive evidence. A gating mechanism compares each query against a patient-level summary of their record and activates only the queries whose evidence type is actually reflected in that patient. Active queries retrieve their top-matching chunks, pool them by attention, and pass them through per-query expert networks. The aggregated representation drives a risk prediction. After training, Integrated Gradients is applied to produce a signed attribution score for each retrieved chunk—quantifying how much it pushed the prediction up or down. The top-scoring chunks become the "evidence pack."

Layer 2 — Rationale Generation. An LLM receives the evidence pack and the prediction probability, and writes a structured rationale. For each factor it produces four fields: a summary, a verbatim quote with a citation ID, brief reasoning over that quote, and the attribution score. A final section synthesizes the factors. Because every quote is checked by string matching against the cited source, the rationale is auditable without reading the whole record.

Layer 3 — Process Verification. A Llama-3.1-8B-Instruct model is fine-tuned on synthetic negative examples—GPT-4o-mini was prompted to corrupt individual reasoning steps in one of five error categories. The verifier outputs per-step probabilities over "correct" and each error type, which are Platt-calibrated at inference so that low-confidence steps get flagged with a predicted error type for clinician attention.

Evaluation. Prediction is measured by accuracy and AUC. Faithfulness is measured by splitting the chain from input to reasoning into two checks: does the quoted observation actually exist in the cited EHR span (string-based longest-common-subsequence matching), and does the reasoning follow from that quote without introducing outside facts (an LLM judge sees only the rationale, not the full record). Each rationale has 22 checks total across five factors and two overall criteria.

Why This Matters

This paper reframes the clinical LLM problem: instead of asking how to fit a patient's whole history into a model, it asks how to build a compact, prediction-optimized scaffold that makes the model's reasoning inspectable. Traceability is not a cosmetic feature here—it is the mechanism that makes clinician trust in a high-stakes setting possible, since every claim can be checked against a specific quoted span rather than against thousands of tokens.

Impact on research: The work shows that retrieval objectives matter more than retrieval volume. Training queries against outcome labels rather than textual similarity consistently outperforms standard RAG, and the finding that attribution-ranked evidence benefits zero-shot LLMs (not just the predictor that produced it) suggests the ranking surfaces generalizable clinical signal. The 2×2 ablation also cleanly separates the effect of better evidence from the effect of a better output format—a distinction much prior work conflates.

Real-world applications:

• Pediatric developmental screening. Early ASD and ADHD prediction from well-child visit histories, well before typical diagnosis age, potentially shortening time to evaluation and support.
• Risk stratification at discharge. One-year all-cause mortality prediction to guide post-discharge follow-up intensity.
• Clinical documentation auditing. The step-level verifier generalizes to flagging unreliable reasoning in any generated clinical narrative, not just EviGen's own output.
• Prior authorization and care-gap justification. Citation-grounded rationales that trace to specific record spans map naturally onto documentation requirements that demand evidence provenance.

Industry relevance: The compute profile is a practical selling point. EviGen's cost is independent of record length and its training runs fit on a single mid-range GPU, whereas full-context fine-tuning requires tens of GPU-hours on top-tier hardware. For health systems with long patient histories, this makes the approach deployable at scale. The architecture is also modular—any LLM backbone can be swapped in—so it can ride improvements in base models rather than requiring retraining.

Future Directions

• Extending to additional EHR modalities. The current system handles notes and ICD codes; lab values, medication histories, and clinical imaging could be added to ground rationales in a fuller view of the patient record.

• Integrating the verifier into the generation loop. The verifier currently runs as a post-hoc check. Feeding its flags back during generation could prevent unreliable steps from being produced at all, rather than catching them afterward.

• A properly powered clinical evaluation. The usability pilot used 7 medical students and is not statistically powered. A larger study with licensed practicing clinicians is described as underway, and would test whether the trust advantage holds under real verification conditions—reviewers in the pilot had to assume predictions were correct and quotes were not hallucinated.

• Robustness in rare-disease settings. EviGen depends on labeled outcomes to train its query vectors. For rare diseases with few positive cases, learning useful queries may be difficult, and the advantage over zero-shot baselines may shrink. Methods for label-efficient query training remain open.

• Subgroup bias monitoring. Because models learn from EHRs, they inherit documentation and access biases—children with sparser records may be systematically missed. Deployment would require monitoring performance across subgroups rather than relying on aggregate accuracy.

• Guarding against automation bias. Fluent, well-cited rationales can be over-trusted even when wrong. The verifier mitigates this but is imperfect and offers no correctness guarantee.

Target Audience

This paper is most valuable to clinical NLP and medical AI researchers working on retrieval-augmented generation, LLM grounding, or EHR-based prediction. It also speaks directly to applied machine learning engineers building decision-support tools in healthcare, who will care about the compute profile and the swappable-backbone design. Clinician-informaticists and health system leaders evaluating AI for clinical deployment will find the faithfulness evaluation design and the honest treatment of automation bias and subgroup fairness relevant to procurement and governance decisions. Readers with a general ML background should be able to follow the framework, though the attribution and gating details benefit from some familiarity with retrieval architectures.

Authors’ abstract

Longitudinal electronic health records (EHRs) capture years of patient history across notes, codes, labs, and procedures, and contain evidence needed to reason about likely clinical outcomes. However, comprehensive clinician review of these records is impractical, and LLM-based processing is costly and often unreliable, missing some relevant observations while hallucinating others. We therefore propose EviGen, a three-layer framework for verifiable clinical rationale generation that addresses these challenges. The first layer is a patient-conditioned retriever that uses learnable queries to find evidence predictive of, not just textually relevant to, a clinical outcome and ranks it by prediction attribution scores. The second layer is an LLM generator that consumes this ranked evidence as a scaffold to produce a clinical rationale grounded in the retrieved spans. The third layer is a process-supervised verifier that checks the generated rationale at the reasoning-step level, flagging unreliable claims. Across three medical prediction datasets, EviGen improves prediction performance and rationale faithfulness over full-context LLM and RAG baselines, and is preferred by clinical reviewers in a usability evaluation.