💻 computer science

Stratum-eval: A Normative-First Evaluation Framework for Clinical Machine Learning

The paper introduces Stratum-eval, a normative-first evaluation framework for clinical machine learning that mandates a validated Normative Specification Document defining use cases, fairness trade-offs, and stakeholder boundaries before computing any performance metrics, thereby ensuring ethical and regulatory alignment is established prior to model deployment.

Original authors: Hassan Farooq, Abdullah Jawad, Muhammad Salman Butt

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hassan Farooq, Abdullah Jawad, Muhammad Salman Butt

Original paper licensed under CC BY 4.0 (https://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 doctor about to use a new, high-tech robot assistant to help diagnose patients. Before you let this robot touch a single patient, you want to know: Is it smart? Is it fair? And most importantly, what kind of mistakes is it allowed to make?

Usually, when we test these medical AI robots, we just ask, "How often is it right?" (like a test score). If it gets 90% right, we say, "Great, let's use it!"

The paper you shared, "Stratum-eval," argues that this is like buying a car without checking if the brakes work or if the steering wheel is on the correct side. It says we are asking the wrong questions first.

Here is the paper's idea, broken down into simple analogies:

1. The "Rulebook" Must Come Before the "Test"

The authors say we are doing things backward. Usually, we build the robot, test it, and then argue about whether the results are fair.

Stratum-eval flips this. Before you even run a single test, you must write a Normative Specification Document (NSD). Think of this as a contract or a rulebook that you sign before the game starts.

What's in the contract? It defines exactly what the robot is for, who might get hurt if it makes a mistake, and—crucially—what kind of trade-offs are acceptable.
The "Hard Stop": If you don't have this signed contract, or if the contract says "we don't care who gets hurt," the system refuses to run. It's like a safety inspector refusing to let a plane take off because the pilot didn't sign the pre-flight checklist.

2. The "Eight Ghosts" (Common Mistakes)

The paper identifies eight ways medical AI evaluations can fail, which they call "failure modes." They are like ghosts that haunt the data, making the robot look good when it's actually dangerous.

The "Echo Chamber" Ghost: The robot is tested on the same people who wrote the answers. It's like a student taking a test where the teacher is also the one who wrote the answer key. The robot just memorizes the teacher's notes instead of learning medicine.
The "Magic Number" Ghost: The robot gives a single score (like "85% accuracy"), but that number hides the fact that it is terrible at helping a specific group of people (like women or older patients).
The "Impossible Promise" Ghost: The paper highlights a mathematical law (the Chouldechova impossibility theorem). Imagine trying to promise that a robot will be 100% fair to two groups of people who have different disease rates. The math says you can't have it all. You have to choose: do you want to miss fewer sick people, or do you want to scare fewer healthy people? You can't do both perfectly. The framework forces you to admit this impossibility before you start.

3. The Five-Layer Inspection

Once the contract (NSD) is signed, the framework runs a five-layer inspection on the robot. Think of this like a car inspection that goes deeper than just checking the engine:

Layer 1 (The Engine): Does the robot actually know the difference between sick and healthy? (Discrimination).
Layer 2 (The Gauge): When the robot says "80% chance of sickness," is it actually 80%? Or is it guessing? (Calibration).
Layer 3 (The Fairness Check): Does the robot make the same number of mistakes for men and women? If the contract said "no more than 15% missed cases for men," does it pass?
Layer 4 (The Intersection Check): What about a specific group, like "older women"? The framework checks if the robot fails them specifically, even if it works fine for everyone else.
Layer 5 (The Time Travel Check): If the robot was trained on data from last year, will it still work next year? Or will the world change and the robot become useless?

4. The "Expiration Date"

This is the most unique part. The authors say that ethical rules change. What was acceptable today might not be acceptable in two years.

So, every NSD contract has an expiration date (a "sunset condition").

If the date passes, the evaluation report becomes invalid.
You cannot just reuse an old report. You have to go back, talk to the stakeholders (patients, doctors), and sign a new contract.
It's like a food label: if the expiration date is past, you throw it away. You don't eat it just because it looked good yesterday.

5. The Real-World Test (The Sepsis Example)

The authors tested their system on a small dataset of ICU patients (people with severe infections).

They found that the robot was actually better at helping women than men.
Because of the math (the "Impossible Promise"), the robot had to miss more sick men to avoid scaring too many healthy women.
The Result: The framework caught this immediately. It didn't just say "The robot is 90% accurate." It said, "Stop! The robot is violating the contract for men. You need to decide: is it okay to miss more sick men, or do you need to change the robot?"

The Bottom Line

The paper isn't saying "AI is bad." It's saying, "We are too lazy about how we check AI."

Currently, we let the engineers build the robot, then we check the score, and then we hope for the best. Stratum-eval says: "No. First, we must sit down and agree on the rules of the game, admit what we can't fix, and sign a contract. If we can't do that, we don't play."

It turns the evaluation from a simple math test into a governance process, ensuring that the people who will be affected by the robot have a say in the rules before the robot ever sees a patient.

Technical Summary: Stratum-eval

Problem Statement

Current clinical machine learning (ML) evaluation frameworks suffer from a critical ordering error: they prioritize discriminative performance metrics while deferring normative questions regarding which errors matter, to whom, and under what conditions a model should be deployed. This approach leads to "latent failures"—plausible-looking metrics that conceal systemic issues such as spurious invariance, fairness impossibilities, and sociotechnical emergence. The paper argues that without a prior commitment to normative specifications, metrics cannot reliably inform clinical deployment decisions. Existing benchmarks often fail to detect these issues because they lack a mechanism to enforce the compatibility of fairness criteria with observed data distributions (e.g., the Chouldechova impossibility theorem) before modeling begins.

Methodology

The authors propose stratum-eval, an open-source evaluation framework that inverts the standard pipeline by making Normative Specification a prerequisite for any metric computation. The methodology consists of three core components:

1. The Normative Specification Document (NSD)

Before any data is touched, evaluators must complete an NSD, a machine-validated artifact that encodes:

Use Case: Plain language description of the model's decision role.
Fairness Criterion: Selection of a specific moral priority (e.g., equalized odds, calibration, demographic parity).
Fairness Rationale: A mandatory statement acknowledging the trade-offs and costs to groups not favored by the chosen criterion.
Acceptable Error Thresholds: Socially negotiated limits for False Negative Rates (FNR) and False Positive Rates (FPR).
Validity Horizon & Sunset Condition: Time-bounded validity and specific conditions for model withdrawal.

The NSD undergoes "hard validation." If fields are missing, the rationale lacks trade-off language, or the validity horizon has expired, the pipeline raises a StratumNSDError and halts.

2. The Five-Layer Evaluation Protocol

Once the NSD is validated, the framework executes a sequential five-layer protocol:

Layer 1 (Invariance-Tested Benchmarking): Detects spurious correlations using the Spurious Correlation Index (SCI) derived from Invariant Risk Minimization (IRM) and variance-based Risk Extrapolation (v-REx).
Layer 2 (Label Quality & Uncertainty): Applies the Dawid-Skene model for label quality (requiring $\ge$ 3 annotators) and checks conformal exchangeability assumptions.
Layer 3 (Group Fairness): Calculates FNR and FPR per group against NSD thresholds.
Layer 4 (Intersectional Disparity): Assesses disparities at the intersection of grouping variables. Cells with $n < 30$ are flagged as underpowered rather than suppressed, retaining visibility while withholding point estimates.
Layer 5 (Temporal Robustness): Computes performance drift across chronological quartiles or uses bootstrap stability checks for non-temporal data.

3. Diagnostic Tools

Fairness Impossibility Visualiser: Pre-computes the Chouldechova impossibility boundary to show whether a chosen fairness criterion is mathematically compatible with observed base rates before model training.
StratumCard: A structured model card generated from the results, containing machine-readable regulatory exports and human-readable summaries anchored to NSD commitments.

Key Contributions

Normative-First Architecture: The framework enforces that normative commitments (fairness criteria, error thresholds) are defined and validated before metric computation, preventing the "post-hoc" rationalization of model failures.
Taxonomy of Evaluation Failures: The paper identifies eight canonical failure modes (F1–F8), categorized by source (Data, Metric, Deployment, Normative) and detectability (Manifest vs. Latent). It highlights that six of these are latent, requiring specific diagnostic tests not found in standard toolkits.
Machine-Enforced Governance: Unlike Model Cards or Datasheets which rely on author good faith, stratum-eval uses code to enforce constraints. An expired NSD prevents the generation of a StratumCard, and missing trade-off rationales cause pipeline failure.
Operationalizing Impossibility: The framework explicitly surfaces the Chouldechova impossibility theorem (the incompatibility of calibration, equalized odds, and differing base rates) as a structural constraint to be resolved via stakeholder deliberation, not algorithmic tuning.

Experimental Results

The framework was applied to the MIMIC-IV Clinical Database Demo (v2.2), a cohort of 140 ICU stays with a Sepsis-3 outcome (prevalence 41.4%).

Dataset Characteristics: The cohort exhibited a base rate difference of 14.8 percentage points between male (48.1%) and female (33.3%) patients, triggering the Chouldechova impossibility condition.
Model Performance: An XGBoost classifier achieved an overall AUROC of 0.905. However, group-level analysis revealed significant disparities.
Layer 3 Findings: The model violated the NSD's acceptable FNR threshold (0.15) for male patients (Group A), with an observed FNR of 0.429. Female patients (Group B) met the threshold (FNR 0.143).
Layer 2 Findings: Group B showed substantially better calibration (ECE 0.110) compared to Group A (ECE 0.237).
Impossibility Analysis: The framework correctly identified that the base rate difference made simultaneous satisfaction of calibration and equalized odds impossible. The NSD had selected calibration as the criterion; the framework surfaced that this choice structurally forced a higher FPR for the lower-prevalence group (females) or, in this specific threshold configuration, a violation of the FNR threshold for the higher-prevalence group (males).
Limitations: Due to the small sample size ( $N=140$ ), Layer 4 (Intersectional) and Layer 5 (Temporal) results were flagged as underpowered or unreliable, serving only as proof-of-concept.

Significance and Claims

The paper claims that stratum-eval shifts the paradigm of clinical AI evaluation from "documentation as a transparency gesture" to "documentation as an operational gate."

Regulatory Alignment: The NSD mechanism directly operationalizes the FDA's AI/ML Action Plan (specifically Action 3 on good ML practices and Action 5 on real-world monitoring) and addresses EU AI Act transparency requirements.
Governance Requirement: The authors assert that institutions without a governance structure capable of producing a named, signed-off NSD are not ready to deploy clinical risk models under this protocol. This is framed not as a limitation of the framework, but as a necessary finding.
Modesty: The authors explicitly state that stratum-eval is an evaluation framework, not a model development or regulatory adjudication tool. It surfaces information to enable responsible decision-making but does not make deployment decisions itself. The framework acknowledges that its current lexical validation of "trade-off" rationales is a proxy for genuine deliberation and that stakeholder representation is documented but not verified in this proof-of-concept.

The core conclusion is that the "harder test" for clinical ML is not whether a model runs on a dataset, but whether a hospital's governance committee is willing to formally commit to specific error thresholds and fairness trade-offs in a time-bounded document before a single metric is computed.