This AI Doesn't Pretend to Know What It Doesn't Know. That's What Makes It Useful for Alzheimer's.
A doctor sits down with a patient showing early signs of memory loss. The MRI is on file. The genetic test was done last month. But the spinal tap results haven’t come back yet, and the cognitive assessment from the referring clinic uses a different scoring system than what this hospital prefers. The doctor has to make a call with what’s available.
That’s not a hypothetical scenario. It’s the daily reality of Alzheimer’s diagnosis worldwide. And it’s the problem a team at Lausanne University Hospital set out to solve.
Background & Context
More than 55 million people live with Alzheimer’s and other dementias globally, a number projected to nearly triple by 2050. The personal cost is devastating: patients lose memory, independence, and identity over years, while families provide an estimated 82 billion hours of informal care annually. Health systems spend over $1.3 trillion per year on dementia-related care.
Recent approvals of disease-modifying drugs are a genuine breakthrough, but they come with a catch. The treatments work best when given early, before substantial neurodegeneration has occurred. Yet diagnosis often comes late because the data doctors need is scattered across different tests, different clinics, and different formats. Some patients have MRI scans but not spinal fluid tests. Others have genetic screening but no cognitive assessments. Attempting to fill in those gaps by guessing (what statisticians call imputation) can create phantom evidence that leads to overconfident, wrong predictions.
A team led by Christelle Schneuwly Diaz and Oliver Chén at Lausanne University Hospital built a different kind of solution. They call it NITROGEN, and it doesn’t guess.
What the Researchers Did
NITROGEN is a transformer-based model (the same architectural family that powers large language models), but designed for clinical tables rather than text. Its core trick: it never imputes missing values. Instead of filling gaps with synthetic numbers, it uses masked attention to ignore missing data entirely while learning from whatever data happens to be present.
The model combines two attention mechanisms. One looks within a single patient’s record, finding patterns between whatever tests are available. The other looks across patients, comparing one person’s profile to others in the dataset. Together, they let the model learn from incomplete records without ever pretending the gaps don’t exist.
What sets NITROGEN apart is what comes after the prediction. The team introduced a modality-aware uncertainty adjustment. If a highly informative test is missing (say, the APOE genetic test), the model expresses more uncertainty about its prediction. If a less critical test is absent, confidence stays high. Think of it as the mathematical equivalent of a doctor saying, “I think it’s Alzheimer’s, but I’d really like to see the genetic results before I’m sure.”
The researchers trained NITROGEN on 7,858 brain scans from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and tested it on two completely independent cohorts without any retraining: OASIS-3 (2,675 scans) and the Australian AIBL study (1,286 scans).
What They Found
For the simplest task — telling Alzheimer’s patients from cognitively normal individuals — NITROGEN hit an AUC of 0.924 on the ADNI test set. AUC runs from 0.5 (coin flip) to 1.0 (perfect), so that’s strong. More revealing is what happened when the model traveled to new hospitals.
On the Australian AIBL cohort, performance held up well, with AUCs between 0.901 and 0.936 across models. On OASIS-3, which targets preclinical populations with fewer Alzheimer’s cases, performance dropped across the board. But that drop reflected a real shift in the patient population. And NITROGEN’s calibration (how well its confidence matched reality) stayed competitive.
Here’s where the comparison gets interesting:
| Model | ADNI Test AUC | AIBL AUC | OASIS-3 AUC | Calibration Quality |
|---|---|---|---|---|
| LightGBM | 0.935 | 0.916 | 0.719 | Poor (overconfident) |
| XGBoost | 0.933 | 0.918 | 0.713 | Poor (overconfident) |
| NITROGEN | 0.924 | 0.901 | 0.710 | Best (well-calibrated) |
| NAIM | 0.915 | 0.905 | 0.703 | Good |
| Random Forest | 0.916 | 0.913 | 0.700 | Moderate |
Tree-based models like XGBoost and LightGBM squeezed out slightly higher discrimination scores, but they were overconfident. Their predicted probabilities didn’t match real-world outcomes. In a clinical setting, an overconfident wrong answer is far more dangerous than a cautious one. NITROGEN’s probability estimates were the most honest.
The harder task (distinguishing cognitively normal people from those with any impairment, including mild cognitive impairment or MCI) was harder for everyone. AUCs dropped to 0.748 to 0.802 on ADNI and 0.639 to 0.700 on OASIS-3. MCI is a notoriously blurry diagnostic category: a transitional state where some people progress to Alzheimer’s and others don’t. No model could crack it cleanly, but NITROGEN stayed competitive and well-calibrated.
For cognitive score prediction (estimating MMSE and CDR-SB scores from clinical data), NITROGEN achieved correlations of r=0.621 on the ADNI test set and r=0.664 on the external AIBL cohort. Tree-based models showed stronger in-distribution performance, but their advantage shrank or disappeared on external evaluation. NITROGEN’s performance changed the least between training and external cohorts, suggesting it doesn’t overfit to the quirks of any single dataset.
Why It Matters
Alzheimer’s diagnosis today is a patchwork. Different hospitals collect different data. A patient might have a detailed MRI at one clinic, genetic testing at another, and cognitive scores from a third. Those records rarely align perfectly. The conventional fix (filling missing values with statistical guesses) can distort the relationships between features and produce overconfident, misleading predictions.
NITROGEN sidesteps the entire problem. It doesn’t need complete data to work, and it tells you how much to trust its answer based on what’s actually available. That’s the difference between a model that works in a pristine research database and one that works in a real hospital.
The team’s feature attribution analysis also surfaced something biologically meaningful. Across all three cohorts and all seven models tested, one brain region consistently emerged as the most important MRI feature: the temporal pole. That’s the forward tip of the temporal lobe, involved in semantic memory and social cognition. Its cortical thickness was the single most informative structural biomarker after APOE genotype. This wasn’t cherry-picked. It showed up in every model, on every dataset.
| Feature | Importance Rank | Consistent Across All 3 Cohorts? |
|---|---|---|
| APOEε4 genotype | 1st | Yes |
| Temporal pole thickness (bilateral) | 2nd–5th | Yes |
| Age | 6th–8th | Yes |
| Education | 8th–12th | Partial |
| Parahippocampal cortex | 10th–15th | Yes (2 of 3) |
How It Could Change Our Lives
A primary care doctor in a rural clinic sees a 68-year-old patient worried about memory lapses. The clinic has no MRI machine, no genetic testing lab, and the nearest neurologist is a three-hour drive away. The doctor can run a basic cognitive screen and note the patient’s age and education. That’s it.
A conventional model might either refuse to make a prediction or spit out an overconfident verdict based on imputed data. NITROGEN would produce a probability (say, 62% chance of impairment) and attach a large uncertainty estimate because key modalities are missing. The doctor gets actionable information plus an honest confidence level. If the uncertainty is high, that’s the signal to order more tests or refer to a specialist.
For drug developers, the implications are just as significant. Clinical trials for Alzheimer’s therapies have long struggled with patient selection. Recruiting people who are too far along means the drug can’t show benefit. Recruiting people who are too early means there’s nothing to measure yet. An imputation-free model that works across different hospitals could help find the right patients at the right treatment window.
The Bigger Picture
NITROGEN belongs to a broader shift in medical AI. For years, the field obsessed over accuracy on clean benchmark datasets. But the hardest problem in clinical machine learning isn’t getting the right answer on perfect data. It’s knowing when you don’t have enough information to give a confident answer at all.
The fact that NITROGEN’s calibration held up better than its raw discrimination scores is, in some ways, the most encouraging result. A model that’s slightly less accurate but honest about its uncertainty is far more deployable than one that’s slightly more accurate and dangerously overconfident.
Limitations & What’s Next
NITROGEN wasn’t perfect. Performance on the OASIS-3 cohort was notably weaker, and MCI classification remained hard for every model. The OASIS-3 challenge reflects a fundamental problem: models trained on clinically diagnosed populations may not transfer well to preclinical or community-based samples where pathology is present but symptoms haven’t fully emerged.
The modality-aware uncertainty adjustment needs broader validation (it was tested with a single random seed). The model currently handles only structured clinical data; integrating raw imaging or free-text clinical notes is an obvious next step. The code is publicly available at github.com/cschneuw/nitrogen. For a field where 55 million people wait for earlier, more reliable diagnosis, even an incremental step toward honest, uncertainty-aware AI matters.
📄 Source: Schneuwly Diaz et al., “Imputation-free transformer learning enables robust Alzheimer’s disease prediction and calibrated uncertainty quantification across heterogeneous clinical cohorts,” arXiv:2607.11656v2, 2026.