Standardizing Longitudinal Radiology Report

Key Technique

The pipeline operates in two sequential stages. Stage 1 (Longitudinal Sentence Detection) takes each sentence from a radiology report and classifies it as either longitudinal (containing a comparison to a prior study) or cross-sectional (describing only current findings). This is a binary classification task where the LLM returns a structured label (1 for longitudinal, 0 for cross-sectional). Sentences like "Pleural effusion has decreased compared to the prior exam" are longitudinal; "No acute cardiopulmonary abnormality" is cross-sectional.

Stage 2 (Disease Progression Extraction) processes only the sentences flagged as longitudinal and maps each to a curated vocabulary of ~50 radiological findings (atelectasis, pleural effusion, pneumonia, cardiomegaly, edema, pneumothorax, etc.), assigning each finding one of four progression labels: improved (finding resolved or decreased), no change (stable), worsened (increased severity or new development), or unmentioned (finding not discussed, used for hallucination detection in generated reports). This two-stage approach avoids running expensive disease-level extraction on irrelevant sentences, and the structured output enables direct F1-score comparison between ground-truth and generated reports.

The critical insight is that medium-scale LLMs (Qwen2.5-32B at ~32B parameters) outperform both larger models and rule-based systems for this task. Larger models (70B+) showed higher recall on sentence detection but lower precision on progression classification due to over-generation. The 32B model hit the optimal cost-accuracy-speed tradeoff at ~2 seconds per query, making corpus-scale annotation (95K+ reports) practical.

Step-by-Step Workflow

1. Define the disease vocabulary. Create a curated list of radiological findings relevant to the domain — typically 30-50 terms for chest X-ray reports (e.g., atelectasis, pleural_effusion, pneumonia, cardiomegaly, pulmonary_edema, pneumothorax, consolidation, lung_opacity, support_devices). Store this as a JSON or YAML config so it can be updated per domain.

2. Preprocess reports into sentences. Split each radiology report's FINDINGS and IMPRESSION sections into individual sentences using a sentence tokenizer (e.g., nltk.sent_tokenize or spacy). Strip section headers and formatting artifacts. Preserve the report-level ID and sentence index for reassembly.

3. Build the Stage 1 prompt (Longitudinal Sentence Detection). Construct a few-shot prompt that defines the task, provides 3-5 positive and negative examples, and instructs the LLM to return a structured binary label. The prompt should emphasize that longitudinal sentences explicitly or implicitly compare current findings to prior studies.

4. Run Stage 1 classification over all sentences. Send each sentence (or batches of sentences) through the LLM with the Stage 1 prompt. Parse the structured output to partition sentences into longitudinal and cross-sectional sets.

5. Build the Stage 2 prompt (Disease Progression Extraction). For longitudinal sentences only, construct a prompt that provides the disease vocabulary and asks the LLM to return, for each mentioned finding, the progression label (improved, no_change, worsened). Use a structured output format (JSON preferred) to enforce schema compliance.

6. Run Stage 2 extraction. Process each longitudinal sentence through the LLM with the Stage 2 prompt. Parse the JSON output into a per-report, per-disease progression matrix.

7. Aggregate per-report annotations. Merge sentence-level extractions into a report-level structure: for each report, produce a dictionary mapping each disease to its progression status, with unmentioned as the default for diseases not referenced.

8. Validate with gold-standard data. If ground-truth annotations exist (or a manual sample is created), compute per-class precision, recall, and F1-score for both Stage 1 (binary) and Stage 2 (multi-class per disease). Use micro-averaged F1 as the primary metric.

9. Benchmark report generation models. To evaluate a generation model, run the same two-stage pipeline on both ground-truth and generated reports, then compare the per-disease progression matrices using F1-score. This directly measures whether the model captures longitudinal clinical information, unlike surface-level metrics (BLEU, ROUGE).

10. Persist annotations in a structured format. Save the annotated dataset as JSONL with fields: report_id, sentence_index, sentence_text, is_longitudinal, disease_progressions (dict). This enables downstream filtering, benchmarking, and model training.

Concrete Examples

Example 1: Annotating a single radiology report

User: "I have a chest X-ray report and I want to extract which diseases improved, worsened, or stayed the same compared to the prior exam."

Approach:

1. Split the report into sentences
2. Classify each sentence as longitudinal or cross-sectional
3. Extract disease progression from longitudinal sentences

# Input report (FINDINGS section)
report = """
Heart size is mildly enlarged, stable compared to prior exam.
There is improved aeration at the left lung base with decreased
atelectasis. Small bilateral pleural effusions persist unchanged.
No pneumothorax. Endotracheal tube is in satisfactory position.
"""

# After Stage 1 (sentence detection):
annotations = [
    {"sentence": "Heart size is mildly enlarged, stable compared to prior exam.",
     "is_longitudinal": True},
    {"sentence": "There is improved aeration at the left lung base with decreased atelectasis.",
     "is_longitudinal": True},
    {"sentence": "Small bilateral pleural effusions persist unchanged.",
     "is_longitudinal": True},
    {"sentence": "No pneumothorax.",
     "is_longitudinal": False},
    {"sentence": "Endotracheal tube is in satisfactory position.",
     "is_longitudinal": False},
]

# After Stage 2 (disease progression extraction):
progression = {
    "cardiomegaly": "no_change",
    "atelectasis": "improved",
    "pleural_effusion": "no_change",
    "pneumothorax": "unmentioned",  # negated, not a longitudinal comparison
}

Example 2: Building the full annotation pipeline script

User: "I need to annotate 10,000 MIMIC-CXR reports with longitudinal labels using a local LLM. Build me the pipeline."

Approach:

1. Set up batch processing with rate limiting
2. Implement both prompt stages
3. Output structured JSONL

import json
from pathlib import Path

DISEASE_VOCAB = [
    "atelectasis", "cardiomegaly", "consolidation", "edema",
    "enlarged_cardiomediastinum", "fracture", "lung_lesion",
    "lung_opacity", "pleural_effusion", "pleural_other",
    "pneumonia", "pneumothorax", "support_devices"
]

STAGE1_PROMPT = """You are a radiology NLP specialist. Determine whether
the following sentence from a chest X-ray report contains longitudinal
information — i.e., it compares the current finding to a prior study.

Return ONLY a JSON object: {"is_longitudinal": true} or {"is_longitudinal": false}

Examples:
- "Cardiac silhouette is stable." -> {"is_longitudinal": true}
- "No acute cardiopulmonary process." -> {"is_longitudinal": false}
- "Pleural effusion has increased since prior exam." -> {"is_longitudinal": true}
- "Lungs are clear." -> {"is_longitudinal": false}

Sentence: {sentence}"""

STAGE2_PROMPT = """You are a radiology NLP specialist. Given a sentence that
contains longitudinal information from a chest X-ray report, extract which
diseases are mentioned and their progression status.

Disease vocabulary: {vocab}
Progression labels: improved, no_change, worsened

Return ONLY a JSON object mapping disease names to progression labels.
Only include diseases explicitly mentioned in the sentence.

Sentence: {sentence}"""


def annotate_report(report_text: str, llm_client) -> dict:
    sentences = split_into_sentences(report_text)
    results = []
    for sent in sentences:
        # Stage 1
        s1_resp = llm_client.query(STAGE1_PROMPT.format(sentence=sent))
        s1 = json.loads(s1_resp)
        entry = {"sentence": sent, "is_longitudinal": s1["is_longitudinal"]}
        # Stage 2 (only if longitudinal)
        if s1["is_longitudinal"]:
            s2_resp = llm_client.query(STAGE2_PROMPT.format(
                sentence=sent, vocab=", ".join(DISEASE_VOCAB)))
            entry["progressions"] = json.loads(s2_resp)
        results.append(entry)

    # Aggregate report-level progression
    report_progression = {d: "unmentioned" for d in DISEASE_VOCAB}
    for entry in results:
        for disease, status in entry.get("progressions", {}).items():
            report_progression[disease] = status
    return {"sentences": results, "report_progression": report_progression}

Example 3: Benchmarking a report generation model

User: "I have ground-truth reports and model-generated reports. How do I evaluate if the model captures longitudinal information correctly?"

Approach:

1. Run the annotation pipeline on both ground-truth and generated reports
2. Compare per-disease progression labels
3. Compute class-level and micro-averaged F1

from sklearn.metrics import classification_report

def evaluate_longitudinal(gt_reports, gen_reports, annotator):
    all_gt_labels, all_gen_labels = [], []
    for gt, gen in zip(gt_reports, gen_reports):
        gt_ann = annotator(gt)["report_progression"]
        gen_ann = annotator(gen)["report_progression"]
        for disease in DISEASE_VOCAB:
            all_gt_labels.append(gt_ann[disease])
            all_gen_labels.append(gen_ann[disease])

    print(classification_report(
        all_gt_labels, all_gen_labels,
        labels=["improved", "no_change", "worsened", "unmentioned"],
        digits=3
    ))

# Output:
#                precision  recall  f1-score  support
#     improved      0.412   0.389    0.400      312
#    no_change      0.634   0.701    0.666     1847
#     worsened      0.298   0.256    0.275      198
#  unmentioned      0.951   0.943    0.947     8643
#    micro avg      0.891   0.891    0.891    11000

Best Practices

• Do: Use few-shot prompting with 3-5 diverse examples per stage — include edge cases like implicit comparisons ("persistent effusion") and negated findings ("no new infiltrate")
• Do: Enforce structured JSON output via system prompts or constrained decoding to prevent free-text drift that breaks downstream parsing
• Do: Use a curated, domain-specific disease vocabulary rather than letting the LLM invent terms — this ensures consistent label spaces across ground-truth and generated reports
• Do: Process Stage 1 before Stage 2 to avoid wasting inference on cross-sectional sentences, which typically constitute 60-70% of report content
• Avoid: Using models larger than necessary — 32B-parameter models match or exceed 70B models on this task while running 3-4x faster; test your target model before committing to corpus-scale runs
• Avoid: Treating unmentioned and improved/no_change/worsened as balanced classes — the class distribution is heavily skewed toward unmentioned, so always report per-class metrics, not just accuracy

Error Handling

Limitations

• Domain specificity. The disease vocabulary and prompt examples are tuned for chest X-ray reports. Adapting to other modalities (MRI, CT, ultrasound) or body regions requires rebuilding the vocabulary and re-validating with a new gold-standard sample.
• Implicit temporal references. Sentences like "chronic cardiomegaly" imply stability without explicit comparison — LLMs may miss these or misclassify them. Accuracy on implicit longitudinal sentences is consistently lower than on explicit ones.
• No severity grading. The pipeline classifies direction of change (better/same/worse) but does not quantify magnitude. "Slightly improved" and "markedly improved" both map to improved.
• Single-report scope. The pipeline annotates individual reports; it does not align findings across a patient's full longitudinal series. Cross-report entity linking is a separate unsolved problem.
• LLM hallucination in Stage 2. The model may extract diseases not actually mentioned in the sentence. Always validate Stage 2 output against the original sentence text, especially for low-frequency findings.
• Language and dataset bias. Validated only on English-language MIMIC-CXR reports from a single US institution. Performance on reports from other languages, institutions, or documentation styles is unknown.

Reference

Wang, X., Figueredo, G., Li, R., & Chen, X. (2026). Standardizing Longitudinal Radiology Report Evaluation via Large Language Model Annotation. arXiv:2601.16753v1. https://arxiv.org/abs/2601.16753v1

Key takeaway: A two-stage LLM pipeline (sentence detection then disease progression extraction) using a ~32B parameter model outperforms both larger LLMs and rule-based lexicon methods for annotating temporal changes in radiology reports, enabling standardized benchmarking of report generation models.3d:["$","$L47",null,{"content":"$48","frontMatter":{"name":"standardizing-longitudinal-radiology-report","description":"Build LLM-based pipelines that automatically detect and classify longitudinal (temporal) changes in radiology reports. Use when the user mentions 'radiology report annotation', 'longitudinal report evaluation', 'temporal change detection in medical text', 'disease progression extraction', 'radiology NLP pipeline', or 'benchmark radiology report generation'."}}]