🤖 AI

Hybrid Decision Making via Conformal VLM-generated Guidance

This paper introduces ConfGuide, a novel hybrid decision-making framework that leverages conformal risk control to generate succinct, targeted textual guidance for human decision-makers, thereby improving decision quality and reducing cognitive load in complex multi-label medical diagnosis tasks.

Original authors: Debodeep Banerjee, Burcu Sayin, Stefano Teso, Andrea Passerini

Published 2026-04-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Debodeep Banerjee, Burcu Sayin, Stefano Teso, Andrea Passerini

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 a senior doctor in a busy emergency room. You have a patient with a chest X-ray, and you need to figure out what's wrong. You know that missing a serious condition (like a collapsed lung) is dangerous, but you also don't want to panic and treat for every minor issue that might be there.

In the past, AI tried to do this job for you, or it gave you a giant, overwhelming list of 20 possible diseases, saying, "It could be any of these!" This is like a GPS that tells you, "You might be in New York, or maybe London, or perhaps Tokyo," leaving you confused and unable to make a decision.

This paper introduces a new system called CONFGUIDE that acts more like a smart, cautious assistant rather than a bossy robot. Here is how it works, broken down into simple steps:

1. The "Safety Net" (Conformal Risk Control)

First, the system looks at the X-ray with a standard AI model. But instead of just guessing, it puts on a "safety net."

Think of this like a fisherman using a net with a specific hole size.

If the net has tiny holes, it catches everything, including the seaweed (false alarms).
If the net has huge holes, the big fish (serious diseases) might slip through.
CONFGUIDE adjusts the net size automatically. It guarantees that no serious fish will ever slip through the holes. It might catch a few extra pieces of seaweed, but it promises: "I will never miss a real disease."

This creates a shortlist of "suspects" (possible diseases) that the doctor must consider.

2. The "Devil's Advocate" (The Vision-Language Model)

Now, the system has a shortlist of suspects. In the old days, it would just hand you the list. But CONFGUIDE does something smarter. It hires a super-smart, unbiased medical detective (a specialized AI) to look at the X-ray and the shortlist.

This detective doesn't just say, "It's Atelectasis." Instead, it writes a balanced report for each suspect:

The "For" Argument: "Here is exactly what I see in the image that supports this disease." (e.g., "I see a dark patch in the right lung.")
The "Against" Argument: "Here is why this might not be the disease." (e.g., "But the shape looks more like fluid than a collapsed lung, and there are no air bubbles.")

It's like having a lawyer for both sides in a courtroom. The AI presents the evidence for the diagnosis and the evidence against it, all in plain English.

3. The Human in the Loop

Finally, the real doctor gets the X-ray, the shortlist, and this balanced report.

The doctor isn't forced to accept the AI's answer.
The doctor isn't overwhelmed by a list of 50 possibilities.
The doctor can read the "For" and "Against" arguments, look at the X-ray, and make the final decision themselves.

Why is this a big deal?

It reduces "Analysis Paralysis": Doctors don't have to guess which of 20 diseases is real. The AI narrows it down to the most likely ones.
It prevents "Anchoring Bias": Sometimes, if an AI says "This is definitely a broken bone," doctors stop looking for other things. By showing the "Against" arguments, CONFGUIDE forces the doctor to think critically: "Hmm, the AI says it's a fracture, but it also admits the evidence is weak. Let me look closer."
It's Safe: The system is mathematically guaranteed not to miss a critical disease. It's better to have a few false alarms than to miss a life-threatening condition.

The Bottom Line

CONFGUIDE is a hybrid team player. It uses math to ensure safety (no missed diseases) and AI to write a clear, balanced summary of the evidence. It doesn't replace the doctor; it gives the doctor a super-powered briefing so they can make the best possible decision for the patient.

In the paper's tests, this approach helped simulated doctors make better decisions than if they just looked at the X-ray alone or just looked at a raw list of AI guesses. It turns the AI from a "black box" that spits out answers into a "transparent assistant" that explains its reasoning.

1. Problem Statement

In high-stakes domains like healthcare, fully automated AI decision-making is often unsuitable due to the potential cost of errors and the need for contextual human judgment. While Hybrid Decision Making (HDM) aims to combine human and machine strengths, existing approaches face two primary limitations:

Separation of Responsibilities: Many methods assign the decision solely to either the human or the AI, limiting true collaboration on the same instance.
Information Overload & Ambiguity: Existing "Learning to Guide" (LtG) frameworks (e.g., SLOG, MEDGELLAN) often provide guidance that compounds information about all possible outcomes. This results in guidance that is difficult to digest, lacks clear argumentative structure, and may leave users uncertain about how to act. Furthermore, standard AI predictions risk missing critical pathologies (False Negatives), which is unacceptable in medical diagnosis.

2. Methodology: CONFGUIDE

The authors propose CONFGUIDE, a novel, fine-tuning-free pipeline designed to generate succinct, targeted, and risk-controlled guidance for human decision-makers. The system operates in three sequential steps:

A. Conformal Risk Control (CRC) for Prediction Sets

Instead of outputting a single label or a probability distribution for all classes, the system first uses a pre-trained classifier to generate class scores. It then applies Conformal Risk Control (CRC) to select a subset of candidate pathologies.

Mechanism: CRC automatically calibrates a prediction threshold ( $\lambda$ ) based on a user-specified target risk level ( $\alpha$ ).
Objective: It enforces a formal upper bound on the False Negative Rate (FNR). This ensures that the system does not overlook likely present pathologies, prioritizing safety over precision in the initial filtering stage.
Output: A "Conformal Set" of flagged pathologies (e.g., "Atelectasis" and "Support Devices" are flagged; "Consolidation" is not).

B. Vision-Language Model (VLM) for Argument Generation

The conformal set is passed to a state-of-the-art Vision-Language Model (specifically MedGemma 27B in the experiments).

Task: The VLM receives the original Chest X-ray image and the list of flagged pathologies.
Output: For each flagged pathology, the VLM generates structured, differential reasoning:
- Arguments For: Radiographic evidence supporting the presence of the pathology.
- Arguments Against: Evidence suggesting the pathology might be absent (e.g., mimics, artifacts, or lack of specific signs).
Design Philosophy: This "for-and-against" structure is designed to minimize anchoring bias, preventing clinicians from blindly accepting the AI's flag without critical evaluation.

C. Human-in-the-Loop Decision

The final decision-maker (a physician or a simulated doctor) reviews the original image alongside the structured guidance. The human retains full responsibility for the final diagnosis, using the AI's risk-controlled set and balanced arguments to inform their judgment.

3. Key Contributions

Novel Framework (CONFGUIDE): A fine-tuning-free pipeline that integrates Conformal Risk Control with VLMs to generate structured, instance-specific guidance rather than raw predictions.
Safety Guarantee: By utilizing CRC, the method provides a mathematically guaranteed cap on the False Negative Rate, addressing the critical safety concern of missing diagnoses.
Bias Mitigation: The generation of "arguments against" alongside "arguments for" encourages critical thinking and reduces the risk of the human decision-maker over-relying on automated flags.
Empirical Validation: A preliminary evaluation on a real-world multi-label medical diagnosis task (Chest X-ray pathology classification).

4. Experimental Setup & Results

Dataset: CHEXPERT dataset (202 frontal images for calibration, 500 for testing).
Models:
- Predictor: Pre-trained CheXzero model (Tiu et al., 2022).
- Guidance Generator: Google's MedGemma 27B.
- Simulated Decision Makers: GPT-4o-mini and Qwen3-vl-8B (used to simulate physician performance).
Baselines:
- Standard: Standard thresholding ( $\lambda=0.5$ ).
- CRC: Using the conformal set directly without VLM guidance.
- CRC++: Simulated doctor receives image + conformal set (no VLM arguments).
- CONFGUIDE: Simulated doctor receives image + conformal set + VLM arguments.

Key Findings:

Overall Performance: CONFGUIDE significantly outperformed all baselines. Using GPT-4o-mini as the decision-maker, CONFGUIDE achieved a Micro-F1 score of 50.76, representing a 15.08% improvement over standard thresholding and a 15.08% improvement over CRC alone.
Role of Guidance: Comparing CONFGUIDE against CRC++ (which had the same safety guarantees but no VLM arguments) confirmed that the generated guidance was the primary driver of performance gains.
Per-Class Analysis:
- Improvements were observed in 10 out of 14 pathologies (e.g., significant gains in Fracture, Lung Lesion, and Lung Opacity).
- Performance degraded slightly on rare pathologies (e.g., Pneumothorax, Pleural Other), likely due to limited training data and subtle radiographic features making guidance generation difficult.
Robustness: The system showed consistent gains across different VLM architectures (GPT-4o-mini and Qwen3), suggesting the benefits stem from the guidance structure rather than the specific reasoning capabilities of one model.

5. Significance and Future Work

Paradigm Shift: CONFGUIDE moves HDM away from "AI makes the decision" or "AI gives a list of possibilities" toward "AI provides a structured, safety-guaranteed argumentation framework."
Clinical Relevance: The approach directly addresses the "black box" problem by providing explainable, balanced reasoning and ensures patient safety by mathematically bounding the risk of missed diagnoses.
Limitations & Next Steps: The current study relies on simulated doctors. The authors acknowledge the need for a large-scale human user study with real clinicians to validate the efficacy of the guidance in reducing cognitive load and improving actual diagnostic accuracy. They also plan to investigate improvements in inter-rater agreement among doctors when using the system.

In summary, CONFGUIDE demonstrates that combining statistical safety guarantees (Conformal Risk Control) with the generative reasoning capabilities of VLMs creates a powerful, safe, and effective tool for hybrid human-AI decision-making in critical medical domains.