💻 computer science

Language Models as Interfaces, Not Oracles: A Hybrid LLM-ML System for Pediatric Appendicitis

This paper introduces ClaMPAPP, a hybrid system that leverages large language models as robust interfaces to extract structured features from clinical narratives for a stable XGBoost classifier, demonstrating superior diagnostic performance and safety in pediatric appendicitis triage compared to end-to-end LLM approaches.

Original authors: Soheyl Bateni, Maryam Abdolali

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Soheyl Bateni, Maryam Abdolali

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to diagnose a sick child with a stomach ache. You have two very different tools to help you:

The "Storyteller" (The Large Language Model or LLM): This is like a brilliant, well-read medical student who can read your messy, handwritten notes and understand the story of the patient's pain. However, this student is prone to "hallucinations" (making up facts that sound real but aren't true) and gets confused if you tell the story in a different order. If you ask them to make the final diagnosis, they might guess wrong.
The "Calculator" (The Machine Learning Model): This is like a super-precise, boring math machine. It is incredibly good at crunching numbers (like age, temperature, and blood test results) to calculate the exact risk of appendicitis. But, it can't read a messy story; it only understands neat, organized spreadsheets.

The Problem:
Doctors have messy stories (clinical notes) but need the precise math for a safe diagnosis. If you ask the "Storyteller" to do the math, they might guess. If you try to force the "Calculator" to read the story, it can't understand it.

The Solution: ClaMPAPP
The authors of this paper built a hybrid system called ClaMPAPP. They didn't try to make the "Storyteller" smarter; instead, they gave them a new job.

Think of ClaMPAPP as a highly efficient factory assembly line:

Station 1: The Translator (The LLM as an Interface)
The "Storyteller" (LLM) sits at the first station. Their only job is to read the messy clinical notes and translate them into a neat, organized checklist. They are not allowed to make a diagnosis. They just say, "Okay, the patient is 12 years old, has a fever of 38°C, and has pain in the lower right."
Station 2: The Safety Inspector (The Validator)
Before the checklist moves on, a strict safety inspector checks it. If the "Translator" accidentally wrote that the patient is 200 years old or has a temperature of 500°C, the inspector catches it and marks it as "missing" instead of letting the bad data through. This prevents the system from being tricked by silly mistakes.
Station 3: The Calculator (The XGBoost Model)
The clean, verified checklist is handed to the "Calculator" (a specific type of math model called XGBoost). This machine does the heavy lifting. It looks at the numbers and calculates the risk of appendicitis. Because it's just doing math on verified numbers, it is very reliable and doesn't get confused by the order of the story.
Station 4: The Reporter (The LLM again)
Finally, the "Translator" comes back to take the Calculator's result and write a clear, easy-to-read summary for the doctor.

Why is this better?
The researchers tested this system against using the "Storyteller" alone (asking the LLM to just guess the diagnosis). Here is what they found:

The "Storyteller" alone is shaky: When the researchers shuffled the order of the sentences in the notes (like telling the story backwards), the "Storyteller" got very confused and made many mistakes. They also missed a lot of real cases of appendicitis (false negatives), which is dangerous in an emergency room.
The Hybrid System is steady: Even when the story was shuffled, ClaMPAPP kept working correctly. Because the "Calculator" only looked at the numbers, the order of the words didn't matter.
Safety First: The most important goal in an emergency room is to not miss a sick child. ClaMPAPP missed far fewer cases of appendicitis than the "Storyteller" alone. It was willing to sound the alarm a bit more often (which might mean more tests for healthy kids) rather than risk missing a sick one.

The Bottom Line
The paper argues that we shouldn't treat AI like an "Oracle" (a magical all-knowing voice that gives the final answer). Instead, we should treat it as an "Interface" (a helpful assistant that organizes information).

By letting the AI handle the language and the math handle the decision, ClaMPAPP creates a system that is safer, more reliable, and easier to trust than asking an AI to just "guess" the diagnosis. The authors tested this specifically on children with stomach pain in German hospitals and found that this "Translator + Calculator" approach worked better than using a smart AI to do everything by itself.

Technical Summary: ClaMPAPP – A Hybrid LLM–ML System for Pediatric Appendicitis

Problem Statement

The integration of Artificial Intelligence into clinical decision support for pediatric appendicitis faces a fundamental dichotomy. Supervised Machine Learning (ML) models trained on structured clinical data (e.g., laboratory values, ultrasound measurements) offer robust, deterministic risk prediction but struggle to interface with the unstructured, narrative nature of real-world clinical documentation. Conversely, Large Language Models (LLMs) excel at interpreting free-text narratives but exhibit critical reliability issues when used as standalone diagnostic engines. These issues include sensitivity to prompt phrasing, positional bias (information ordering), and the generation of "hallucinated" but plausible-sounding outputs. In high-stakes acute triage settings, such as pediatric appendicitis, the risk of false negatives (missed diagnoses) is a primary safety concern, making the instability of end-to-end LLM diagnosis unacceptable.

Methodology: The ClaMPAPP Architecture

To address the gap between natural-language usability and diagnostic reliability, the authors propose ClaMPAPP (Clinical Language-assisted Machine-learning Pipeline for Appendicitis). This hybrid system reassigns the LLM from a decision-maker to an interface, delegating the final risk prediction to a supervised ML model. The pipeline consists of four sequential stages:

LLM-Based Feature Extraction: An LLM (specifically Llama-3.1-8B in the rewriting phase) processes unstructured, note-like clinical narratives. Instead of generating a diagnosis, the LLM is constrained via tool calling to perform semantic parsing, mapping free text into a schema-constrained structured representation (JSON) containing demographics, clinical scores (Alvarado, PAS), symptoms, laboratory variables, and ultrasound findings.
Deterministic Feature Validation: A safety gate validates the extracted features against predefined physiological and clinical plausibility constraints (e.g., $0 \le \text{Age} \le 18$ , $35.0 \le \text{Temperature} \le 42.0^\circ\text{C}$ ). Values violating these bounds are mapped to missing values (NaN) rather than being clipped, preventing the introduction of biased inputs. This step ensures that only clinically plausible data enters the predictor.
Risk Prediction: The validated feature vector is passed to an XGBoost classifier trained on structured clinical, laboratory, and ultrasound variables. XGBoost was selected for its ability to handle missing data natively and its strong performance in clinical classification tasks. The model outputs a continuous appendicitis risk score.
Report Generation: The LLM acts as a clinical summarizer, contextualizing the XGBoost risk score and extracted features into a structured report for clinicians, using non-prescriptive triage language (e.g., "High risk: consider surgical consultation") without issuing definitive diagnostic directives.

Data Generation and Evaluation:
To enable controlled evaluation with preserved ground truth, the authors synthesized note-like clinical narratives from structured Electronic Health Record (EHR) data using a template-based rendering followed by LLM rewriting. This approach allowed for the creation of "permuted" narratives where the order of information was swapped to test robustness against positional bias, a known vulnerability of LLMs.

The system was evaluated on two independent retrospective cohorts from German hospitals:

Internal Cohort (Regensburg): 782 patients (580 training, 202 test).
External Cohort (Düsseldorf): 301 patients (used exclusively for external validation).

Key Contributions

The paper outlines three primary contributions:

Repositioning the LLM: The framework constrains the LLM to a feature extraction and interface role rather than direct diagnostic classification. This makes the final prediction deterministic, transparent, and auditable by relying on a validated ML model for risk estimation.
Safety-Oriented Performance: ClaMPAPP demonstrates a superior sensitivity-specificity trade-off compared to standalone LLM baselines. It significantly reduces false negatives (missed appendicitis cases), which is the critical safety metric in acute triage.
Robustness to Narrative Perturbations: Through sentence-order permutation experiments, the study shows that while end-to-end LLM baselines suffer significant performance degradation due to positional bias, the hybrid ClaMPAPP architecture remains comparatively stable.

Experimental Results

The system was benchmarked against a spectrum of LLMs (including OpenAI GPT-5.5, o3-pro, Anthropic Claude Opus 4.7, and open-source models like MedGemma and Llama-3.1) and traditional ML models.

Diagnostic Performance:
- Internal Validation (Regensburg): ClaMPAPP achieved the highest accuracy (85.1%) and F1-score (84.8%) with a sensitivity of 97.7%, missing only 2 confirmed cases. In contrast, standalone LLMs showed unstable trade-offs; for instance, MedGemma-4b-it had high specificity but missed 40 cases (sensitivity 53.5%), while OASST-Pythia-12b missed only 1 case but had a false-positive rate of 113 (specificity 2.6%).
- External Validation (Düsseldorf): ClaMPAPP maintained the highest accuracy (80.7%) and F1-score (0.881) with a sensitivity of 93.5%, missing 15 cases. It outperformed all baselines in minimizing missed diagnoses, whereas models like OpenAI GPT-5.5 missed 78 cases despite higher specificity.
Robustness Analysis:
- Under sentence-order permutation (swapping the first and second halves of the narrative), pure LLM baselines exhibited significant performance collapse. For example, Llama-3.1-8b accuracy dropped to 0.475, and MedGemma-4b-it sensitivity plummeted to 0.128.
- ClaMPAPP demonstrated high architectural robustness, maintaining perfect sensitivity (1.000) internally and 0.926 externally under permutation, with only a minor accuracy decrease (5.2%).
Calibration:
- While ClaMPAPP showed excellent discrimination (Internal AUC = 0.982; External AUC = 0.804), raw probability calibration was imperfect. Exploratory post-hoc logistic recalibration improved calibration metrics (e.g., reducing Brier score and ECE) without altering ranking performance, suggesting the miscalibration was due to probability scale mismatch rather than poor predictive ability.

Significance and Claims

The paper claims that the "LLM-as-interface, ML-as-predictor" paradigm offers a more auditable and safety-oriented pathway for clinical decision support than unconstrained generative diagnosis. By separating natural-language usability from predictive inference, the system leverages the clinical utility of LLMs for data extraction while relying on validated, deterministic ML models for the critical decision step.

The authors emphasize that ClaMPAPP is designed as a triage-oriented risk-stratification aid, not a standalone diagnostic test. Its primary value lies in minimizing missed diagnoses (false negatives) in acute settings. The study concludes that while end-to-end LLMs are currently unsuitable for safety-critical diagnostic decisions due to instability and hallucination risks, hybrid systems can effectively bridge the gap between narrative documentation and structured predictive models. The authors note that this integration pattern is transferable to other medical domains where validated tabular predictors and narrative documentation coexist, though they caution that prospective validation in real-world workflows is necessary before clinical deployment.