1 of 18

AUTHORS: YAQIAO HUANG, PALOMA RABAEY, HENRI ARNO, AND THOMAS DEMEESTER

FROM CLINICAL TEXT TO CAUSAL GRAPHS:�HOW MUCH SUPERVISION IS NEEDED TO INDUCE GRAPH-USABLE SYMPTOM VARIABLES?

HC@AIxIA 2026, Oct. 6 - 8, Perugia, Italy

2 of 18

MOTIVATION

Correlation can be misleading

Lighter ↔ lung cancer: correlated
Smoking → lung cancer: causal
In medicine, we care about causal structure, not only association.

Causation ≠ Correlation

3 of 18

WHY THIS MATTERS IN CLINICAL DATA

Nodes = clinical variables
Edges = possible causal/dependency relations
But many important symptoms are described in free-text notes.

Causal discovery needs graph nodes

Can we convert clinical text into symptom variables that are useful for downstream causal graph learning?

Our task

4 of 18

CORE PROBLEM

Semantic patterns may be coherent
But they may not match the clinical variables we need
Fully labeling patient records is expensive
We need a practical level of supervision

text-derived variables may not be graph-usable

5 of 18

RESEARCH QUESTION

Then three options:

1. Unsupervised: no target guidance
2. Weakly supervised: symptom anchors only
3. Fully supervised: patient-level labels

How much supervision is needed to induce graph-usable symptom variables from clinical text?

6 of 18

TASK SETTING: SIMSUM

10,000 simulated primary-care encounters
Each patient has clinical notes + structured variables

Dataset: SimSUM

Target symptoms:

Cough, Fever, Nasal, Pain, Dyspnea

Figure 2. Adapted SimSUM Bayesian network highlighting the symptom, diagnosis, and outcome variables used in this study.

To study this in a controlled setting, we use SimSUM.

SimSUM contains 10,000 simulated primary-care encounters in the respiratory-disease domain. It is especially useful for this problem because the data are generated from an expert-designed Bayesian network. On the right, you can see an adapted sketch of the gold-model Bayesian network, where I focus on the symptom-related part to make the setup easier to follow. If you're interested in the full original network, you can find it in the appendix.

So for each patient, we have three things: a clinical narrative, structured clinical variables, and the gold symptom variables.

We focus on five symptoms: cough, fever, nasal symptoms, pain, and dyspnea.

Our task is to hide these five symptom variables, recover them from the clinical note, and then see whether the recovered variables allow us to reconstruct the underlying graph.

7 of 18

TASK SETTING

Clinical note:

“The patient reports persistent cough and shortness of breath…”

Recover:

Symptom Value

Cough 1

Dyspnea 1

Fever 0

Nasal 0

Pain 0

→ Use these variables as nodes for downstream graph learning

what we observe and what we recover

8 of 18

GRAPH LEARNER: PC-LEARN

Downstream graph learning

Outline of PC algorithm Image by Shaw Talebi.

9 of 18

BASELINES

Upper reference: G-TAB: using gold symptom variables

Lower baseline: LB-TAB without symptom variables

10 of 18

THREE TEXT-BASED PIPELINES

G-US & G-WS

G-US: Select clusters by quality

Homogeneity → Do members of the cluster describe the same symptom?

Stability → Does the cluster remain similar across repeated runs?

G-WS: Select clusters using symptom anchors

Example: dyspnea & {“shortness of breath”, “difficulty in breathing”, “air hunger”}

Embed anchors in a simple clinical template: “The patient reports [ANCHOR].”

→ Select clusters with the highest semantic similarity to the anchor sentences

We first start with the lowest-supervision setting, G-US.

Here, the clinical notes are split into small clinical statements, embedded, reduced with UMAP, and clustered with HDBSCAN.

We then keep clusters that are relatively homogeneous and stable.

The important thing is that G-US never sees the target symptom names and never sees patient labels. So the data are free to organize themselves.

The question is whether these naturally discovered semantic groups happen to correspond to the clinical variables we need.

G-WS keeps exactly the same unsupervised representation and clustering stages.

But now we introduce a very small amount of domain knowledge. For each symptom, we provide only four or five representative phrases.

For example, for dyspnea we might use phrases such as “shortness of breath”, “difficulty in breathing”, or “air hunger”.

We embed these phrases and use them to identify which existing clusters are semantically related to the target symptom. If several clusters match the same symptom, we merge them into one patient-level variable.

So we are not telling the model which patients have dyspnea. We are only telling it what we mean by dyspnea.

That distinction is really the central idea of G-WS.

11 of 18

THREE TEXT-BASED PIPELINES

G-FS

Vary the amount of labeled training data

5% · 10% · 20% · 50% · 100%

Patient representation

Sentence embeddings → mean pooling → 768-d patient vector

Classification head

768 → 256 → 5 symptom outputs

(two-layer MLP for multi-label classification)

→ Predict the presence/absence of each symptom for every patient

These two settings tell us what happens if symptoms are perfectly observed, or completely missing.

这里需要先对比介绍G-US和G-WS，每个component的，这里的重点是：T5用的heuristics两个指标，homo+stable -> high quality clusters; T6用的anchor具体是什么; T7的2-layer nn model，以及用不同size trainset得到的 head-only classification model；后面还要想一下justify为啥G-FS没做sentence segmentation

这里最好是

Finally, G-FS takes the conventional supervised route.

Instead of discovering clusters, we directly train a multi-label classifier to predict the five symptoms from each patient's text. We mean-pool the sentence embeddings into one patient representation, then pass it through a simple two-layer MLP for symptom prediction.

This gives much stronger task-specific guidance, but now we need patient-level annotations.

We therefore vary the amount of labeled training data from 5% all the way to 100%, so that we can see how graph recovery changes as supervision increases.

12 of 18

RESULTS: VARIABLE INDUCTION

Does each method recover the target symptoms?

13 of 18

RESULTS: GRAPH RECOVERY

Do better symptom variables lead to better graphs?

More importantly, this improvement also appears in the downstream graph.

Without symptoms, graph F1 is 13.3%. G-US barely improves this, reaching 14.6%.

But G-WS jumps to 49%, already much closer to the supervised models. At 20% labeled data, G-FS reaches 57.7%, and full supervision reaches 64.2%.

Put differently, G-WS recovers about 69% of the graph-F1 improvement achieved by full supervision over G-US.

And the supervision cost is very different. Defining the anchors took roughly ten minutes in our timing exercise, while labeling even the smaller supervised subsets would take hours.

So the interesting point is not that weak supervision beats full supervision — it doesn't overall. It's that a small amount of concept-level guidance buys a surprisingly large amount of structural performance.

14 of 18

TRADE-OFF

Comparisons between G-WS and G-FS

15 of 18

WHAT SHOULD WE TAKE AWAY?

Limitations

Synthetic benchmark
Five symptoms only
Anchor / threshold sensitivity
One implementation per supervision regime

Takeaways

Text-derived variables need target alignment
Weak supervision can provide it cheaply
Better variables lead to better graph recovery

16 of 18

Yaqiao Huang

First author & Corresponding author

MASTER OF STATISTICAL DATA ANALYSIS

GHENT UNIVERSITY, BELGIUM

yaqiao.huang@ugent.be

www.ugent.be

Thank You for Listening!

17 of 18

APPENDIX

18 of 18

APPENDIX