Reliable Extraction of Clinical Follow-Up Instructions: A Hybrid Neural-Symbolic Pipeline

arXiv cs.CL Papers

Summary

This paper presents a hybrid neural-symbolic pipeline for extracting follow-up instructions from clinical notes, using BioBERT and deterministic date arithmetic. It achieves high performance (Pair F1 ~0.99) compared to generative baselines.

arXiv:2605.26560v1 Announce Type: new Abstract: Objective. Outpatient notes carry follow-up instructions pairing actions with future times ("MRI brain in two weeks"). Extracting (action, date) pairs supports scheduling and audit, but generative extractors miss the date because linking and arithmetic are implicit in decoding. We test a hybrid neural-symbolic pipeline against direct generation. Methods. We define TestSpecification and TimeSpecification entities and a ScheduledFor relation. BioBERT feeds BIO tagging and a biaffine linker; entities are canonicalized via a 28-action ontology and times normalized to day offsets deterministically. We evaluate on a 2,000-note synthetic outpatient corpus with action-disjoint splits (18 train, 6 OOV-test) against zero-shot GPT-4o-mini and LoRA-fine-tuned LLaMA-3 8B with note-level bootstrap 95% CIs. Results. On 259-note seen and OOV splits the hybrid pipeline achieves Test-Time Pair F1 of 0.997 and 0.986 with 0.00-day MAE. Baselines reach high action F1 (LLaMA-3 0.992; GPT-4o-mini 0.963 seen) but Pair F1 stays at 0.51-0.57 (LLaMA-3) and 0.53 (GPT-4o-mini), CIs non-overlapping with the hybrid. Conclusion. Separating learned entity extraction from deterministic date arithmetic outperforms generation on this benchmark, generalizes to held-out actions, and exposes failure modes. Transfer to real EHR notes is the next validation; a first-pass realism check is in Limitations.
Original Article
View Cached Full Text

Cached at: 05/27/26, 09:07 AM

# Reliable Extraction of Clinical Follow-Up Instructions: A Hybrid Neural-Symbolic Pipeline
Source: [https://arxiv.org/abs/2605.26560](https://arxiv.org/abs/2605.26560)
[View PDF](https://arxiv.org/pdf/2605.26560)

> Abstract:Objective\. Outpatient notes carry follow\-up instructions pairing actions with future times \("MRI brain in two weeks"\)\. Extracting \(action, date\) pairs supports scheduling and audit, but generative extractors miss the date because linking and arithmetic are implicit in decoding\. We test a hybrid neural\-symbolic pipeline against direct generation\. Methods\. We define TestSpecification and TimeSpecification entities and a ScheduledFor relation\. BioBERT feeds BIO tagging and a biaffine linker; entities are canonicalized via a 28\-action ontology and times normalized to day offsets deterministically\. We evaluate on a 2,000\-note synthetic outpatient corpus with action\-disjoint splits \(18 train, 6 OOV\-test\) against zero\-shot GPT\-4o\-mini and LoRA\-fine\-tuned LLaMA\-3 8B with note\-level bootstrap 95% CIs\. Results\. On 259\-note seen and OOV splits the hybrid pipeline achieves Test\-Time Pair F1 of 0\.997 and 0\.986 with 0\.00\-day MAE\. Baselines reach high action F1 \(LLaMA\-3 0\.992; GPT\-4o\-mini 0\.963 seen\) but Pair F1 stays at 0\.51\-0\.57 \(LLaMA\-3\) and 0\.53 \(GPT\-4o\-mini\), CIs non\-overlapping with the hybrid\. Conclusion\. Separating learned entity extraction from deterministic date arithmetic outperforms generation on this benchmark, generalizes to held\-out actions, and exposes failure modes\. Transfer to real EHR notes is the next validation; a first\-pass realism check is in Limitations\.

## Submission history

From: Yehudit Aperstein \[[view email](https://arxiv.org/show-email/0d2b913b/2605.26560)\] **\[v1\]**Tue, 26 May 2026 05:14:33 UTC \(939 KB\)

Similar Articles

Neuro-Symbolic AI for LEED compliance: Document-Centric Benchmarking, Deterministic Numeric Checking, and When Multimodal Hurts

arXiv cs.AI

This paper introduces a neuro-symbolic pipeline for automating LEED v4.1 BD+C compliance verification using small locally deployed language models and deterministic numeric checking. Experiments on four university buildings show that a 4B model outperforms an 8B model, and the deterministic checker corrects arithmetic errors on key credits, though multimodal inputs reduce accuracy.

Training Large Language Models to Predict Clinical Events

arXiv cs.LG

This paper extends Foresight Learning to clinical event prediction by converting time-ordered clinical notes into prediction examples. A LoRA adapter on a 120B model improves calibration and outperforms GPT-5 on held-out questions.