médical IA

Representing each patient as a graph to skip data cleaning: PatTree and its 98.5% accuracy on Alzheimer's classification

Published on August 9, 2026 · 11 min read

Julia Gehrmann, Lars Quakulinski, Hamza Naseem, and Oya Beyan propose PatTree, a way of representing each patient as a knowledge graph built automatically from raw, heterogeneous clinical data, without the long and costly harmonization step that usually precedes analysis. On a subset of the ADNI-1 cohort (763 subjects), classifying patients directly from this graph reaches a balanced accuracy of 98.5% and an F1 of 0.987 to sort Alzheimer's disease, mild cognitive impairment, and normal cognition. The ambition — doing away with manual data cleaning — is real and welcome, but a score this close to perfect on a notoriously hard task, obtained on a single cohort and without external validation, should be read first as a methodological red flag, not as proof.

The context

Before training a model on real medical data, you have to prepare it, and that preparation is the silent bottleneck of health AI. A patient's data are multimodal — of different natures: free-text clinical notes, lab results, imaging, cognitive test scores, demographics. They arrive with missing values, at different dates, in inconsistent formats, and with vocabularies that do not match from one source to the next. Making all of this fit into a single clean table — harmonization — is long, manual, error-prone work that is rarely reproducible from one project to the next.

The consequence is twofold. First, a huge share of a medical AI project's effort goes into cleaning rather than modeling. Second, every harmonization choice — which variable to keep, how to encode a missing value, which format to impose — injects human assumptions that can bias what the model learns. The idea gaining ground in recent years is to represent the patient no longer as a table row but as a knowledge graph: a network of nodes (the data) linked by edges (their relationships), which naturally absorbs heterogeneity and gaps.

This is the context for PatTree, tested here on a classic neurology problem: telling apart, from a patient's data, Alzheimer's disease, mild cognitive impairment (MCI, an intermediate stage where abilities decline without reaching dementia), and normal cognition. The testbed is ADNI (Alzheimer's Disease Neuroimaging Initiative), a large public database of neuroimaging and clinical data widely used by the community.

The method

PatTree automatically structures a patient's raw data into a single knowledge graph. Concretely, each element — a lab value, a test score, a demographic feature — becomes a node, and the relationships between these elements become edges, all connected to the patient. The authors' central claim is that this construction happens without pre-standardization: no common format or fixed table is imposed upstream. The graph preserves the semantic relationships between data of different modalities and sources, making it interoperable (readable and combinable across systems) and machine-interpretable.

The claimed contribution is therefore not a new classifier but a new substrate: instead of cleaning the data to fit a model, you leave them in their diversity and classify directly from the graph. The authors state that classification is "directly feasible" on PatTree, that is, without the usual preparation step.

To evaluate it, they use a subset of 763 ADNI-1 subjects and a three-class task: Alzheimer's, MCI, normal cognition. Two metrics are reported. Balanced accuracy is the average of the correct-detection rates class by class — useful when classes are unequal in size, because it prevents a model from looking good simply by favoring the majority class. The F1-score combines into a single number the precision (of the cases predicted positive, how many truly are) and the recall (of the true positives, how many are found); 1.0 is the maximum. The figures are measured on a held-out test set kept apart from training. The abstract does not specify the comparator used to justify the "state of the art" claim — a point we return to below.

The results

On the test set, PatTree reaches a balanced accuracy of 98.5% and an F1 of 0.987 for the three-class classification. Taken at face value, this means the model is wrong very rarely, and about as rarely for each of the three categories — including MCI, which is precisely the class that usually drags scores down because it overlaps the other two.

This is where reading should slow down. Telling apart Alzheimer's, MCI, and normal cognition is a notoriously hard task: the border between MCI and normal cognition, like the one between MCI and early dementia, is blurry even in clinical diagnosis itself. Near-perfect separation is not the expected result of an intrinsically ambiguous problem. Put in cautious clinical terms: on paper, such a model would miss almost no patient and over-classify almost none; but a score that erases a real difficulty of the field should raise the suspicion that the difficulty was bypassed somewhere in the chain rather than solved.

In other words, the impressive figure is not, in itself, good news: it is an invitation to check where the performance comes from. We are not accusing the authors of cheating — nothing indicates it — but a preprint announcing 98.5% on this task shifts the burden of proof: it must demonstrate that this result is not an artifact.

What is good

A real problem, attacked from the right end. The bottleneck of medical AI is not only the algorithm, it is data preparation. By seeking to remove the harmonization step rather than optimize it, PatTree tackles a cause of non-reproducibility and cost that most papers pass over in silence. The direction is right, and rare.

An interoperable, reusable substrate. Representing the patient in a single knowledge graph that preserves semantic relationships across modalities aligns with the FAIR principles (findable, accessible, interoperable, reusable data) promoted by the medical informatics community. If the graph is genuinely built automatically and without strong assumptions, the idea transfers well beyond Alzheimer's, to any heterogeneous clinical dataset.

Appropriate metrics and a separate test set. Reporting balanced accuracy and F1, rather than raw accuracy alone, is the right choice for a potentially imbalanced multi-class task. Using a held-out test set, rather than mere cross-validation on the training data, is also basic good practice. Since the ADNI data are public, the direction of the work is verifiable by other teams.

What is less good

A near-perfect score that evokes data leakage. Data leakage is the number-one failure mode of the health-AI literature: the model has access, directly or indirectly, to the information it is supposed to predict. Here, two classic mechanisms are plausible and not ruled out by the abstract. On one hand, ADNI contains several time points per patient: if data from the same subject end up in both training and test, performance is inflated. On the other, the diagnostic category (Alzheimer's / MCI / normal) is partly defined by cognitive scores (such as MMSE or CDR); if these scores are in the graph, the model is not predicting the disease, it is re-reading the label. An F1 of 0.987 on this task is exactly the signature one would expect from such leakage.

A single cohort, small sample, no external validation. Seven hundred sixty-three subjects, all from ADNI — a research cohort, selected, not representative of a real patient population. Nothing says the graph built on ADNI would transfer to another hospital, another country, other protocols: this is the classic population bias. Yet PatTree's selling point is precisely generalization to real, heterogeneous data; demonstrating it on a single clean, well-labeled cohort does not test that promise, it sidesteps it.

An unstated comparator and a risk of shortcut learning. The abstract claims a "state of the art" level without naming the baseline against which 98.5% is compared — yet, on an already well-studied task, the value of a figure depends entirely on what it is compared to. Added to this is the risk of shortcut learning: the model may latch onto a shortcut correlated with the diagnosis (a near-definitional test, a collection artifact) rather than the disease signal. Without an ablation study or interpretability analysis showing what the graph relies on, one cannot tell genuine understanding from a shortcut.

What it changes

For the research community, the guiding idea — doing away with manual harmonization by leaving data in a knowledge graph — deserves to be pursued, because it targets a real and underestimated cost. But the bar of proof is clear: the next version will need a data-leakage audit (strict per-patient separation, exclusion of variables near-definitional of the label), external validation on at least a second cohort, and a named comparator. Until those three are present, the spectacular result remains a hypothesis, not a demonstration.

For clinicians, nothing to deploy, and this must be said plainly: classifying ADNI subjects into Alzheimer's / MCI / normal is not a clinical need — diagnosis is made otherwise, and for these subjects it is already known. PatTree's potential value is not this result but the upstream promise: pipelines able to ingest real, messy records without months of preparation. That promise is not yet demonstrated.

For patients and the public, the lesson is a useful reading reflex in the face of any medical-AI announcement: a figure near 100% on a problem that doctors themselves find hard should raise suspicion, not enthusiasm. The best studies often report more modest scores, precisely because they took care not to fool themselves. Performance is worth only as much as the conditions under which it was measured.