Normative Deviation Features for Small-Sample Preclinical Disease Detection

Many diseases, including the neurodegenerative ones, leave traces in scans and blood tests years before anyone is diagnosed. A model that spots those traces in people who still look healthy could move treatment and trial enrollment much earlier. The obstacle is data. In any cohort, the number of people who later convert from healthy to diseased is small, often a few dozen to a few hundred, while each person carries a high-dimensional image or signal plus a spreadsheet of clinical variables. A popular fix for small samples is the "hybrid" classifier. A pretrained neural network turns the image into a feature vector, and a classical statistical model such as a gradient-boosted tree ensemble, a random forest, or a support vector machine makes the final decision in place of the network's own output layer. This note argues that the hybrid recipe, as it is usually applied, puts the split between network and statistical model in the wrong place for preclinical problems.

The core idea is as follows. Before the classical classifier sees anything, express each person's image embedding as a deviation from a model of what healthy people of the same age and sex look like. That model is fitted on the large pool of unlabeled or healthy participants that every cohort has in abundance. The tree ensemble then works with a handful of interpretable "how unusual is this person, along which directions" scores, joined to clinical variables. It never sees hundreds of raw embedding coordinates. Our claim is that this transformation, not the choice of tree versus linear model, is what lets hybrid classifiers work when positives are counted in dozens.

Where the existing hybrid work stops

The motivating evidence comes from a line of work that replaces a network's softmax layer with a classical learner on frozen or pretrained features. Each study runs in a different domain:

Together these establish that decoupling representation from decision helps when labels are scarce. That is the method we reuse.

All five papers hand the classifier the raw embedding, and all of them work in settings where the positive class is not defined by a subtle, age-confounded shift away from normal. That assumption breaks for preclinical detection, for three reasons:

  • Disease is entangled with ageing. In a healthy-looking person the disease signal is a small displacement that is easily confused with normal ageing, sex differences, or scanner effects. Raw embedding coordinates mix all of these together.
  • Tree splits fit raw embeddings poorly. Tree ensembles cut along individual coordinates. Embedding coordinates are arbitrary rotations of the underlying factors, so a tree with a few hundred labeled people and hundreds of coordinates must spend its few splits rediscovering structure that a normative model can extract for free from unlabeled data.
  • Clinical data is messy. Multimodal records have missing modalities and wildly different feature scales. Trees handle both natively, but only if the image side arrives in a comparably compact, calibrated form.

The research file's section on prior work at this exact intersection was truncated. I could not verify whether normative-deviation inputs to hybrid classifiers have been tested on preclinical cohorts. The proposal is therefore framed so that its central comparison stays informative whatever the answer turns out to be.

Core Claim

For preclinical disease identification with few converters, three steps should beat both feeding raw embeddings to the same classifier and fine-tuning end to end:

  1. Fit a normative model of image embeddings on healthy and unlabeled participants, conditioned on age, sex and acquisition site.
  2. Reduce each labeled person's embedding to standardized deviation scores along a small number of directions.
  3. Combine those scores with clinical variables in a gradient-boosted tree ensemble.

The advantage should widen as the number of positives shrinks. It should also hold when training and test sites differ, because the normative step absorbs site and age structure that the classifier would otherwise mistake for disease.

Research Program

The pipeline has three stages, and each is a deliberate design choice that can be ablated.

Encoder. A frozen pretrained encoder maps each scan to an embedding $z \in \mathbb{R}^d$. The encoder is never trained on disease labels.

Normative model. On the healthy reference pool we fit a conditional Gaussian. Its mean $\mu(a)$ and covariance $\Sigma(a)$ depend on covariates $a$ (age, sex, site). The mean is fitted by regression and the covariance is shrunk toward a pooled estimate. We take the top $k$ eigenvectors $u_1,\dots,u_k$ of the pooled covariance and define, for each person,

$$s_j = \frac{u_j^\top (z - \mu(a))}{\sqrt{u_j^\top \Sigma(a)\, u_j}}, \qquad j = 1,\dots,k,$$

so that $s_j$ is a z-score: how many healthy-population standard deviations this person sits from age-matched peers along direction $j$. We add the Mahalanobis norm of the full deviation as one summary score. We will also try a second set of directions taken from a contrastive split, in which healthy-pool variance is compared against variance in the unlabeled pool. These directions can capture where at-risk people spread out more than healthy ones.

Classifier. A gradient-boosted tree ensemble receives the $k+1$ deviation scores, the clinical variables, and indicators for missing modalities. Its job is to predict later conversion. We set $k$ between 4 and 32 and choose it by nested cross-validation.

Two analyses follow from this design. The first is a learning-curve study, which checks whether the benefit concentrates at small positive counts as the claim predicts. The second is an attribution study. Because each $s_j$ has a meaning (a direction of abnormality relative to healthy peers), we can ask which directions the trees use. We can also check whether those directions are stable across bootstrap resamples, which is a property clinicians need before trusting a preclinical flag.

Proposed Experiments

The file contains no benchmark ledger for preclinical multimodal detection, so the datasets below are my proposal and do not come from cited work.

The natural setting is a large population cohort with baseline imaging, baseline clinical and laboratory data, and linked later diagnoses. In such a cohort the healthy majority supplies the normative pool, and incident cases diagnosed years after imaging supply rare positives. UK Biobank brain MRI with incident dementia or Parkinson's disease is one example. Retinal photographs with incident cardiovascular or neurodegenerative outcomes are another. A longitudinal Alzheimer's cohort with cognitively normal participants who later progress gives a second, smaller, independently collected test set.

Training conditions. We subsample training positives to 25, 50, 100 and 200 while keeping the prevalence realistic.

Metrics. We report area under the ROC curve, area under the precision-recall curve (the more honest metric at low prevalence), and calibration error. Every result comes with bootstrap intervals over 20 or more repeated subsamples.

Held-out tests. The same protocol is repeated under leave-one-site-out evaluation and under transfer from the large cohort to the small one.

Baselines

Clinical variables only. The naive reference point is a gradient-boosted tree ensemble on clinical and laboratory variables alone, with no imaging. In preclinical prediction, age and a few blood markers often carry most of the signal. Any imaging pipeline that cannot beat this has not justified the scan. We tune it with the same nested cross-validation and the same hyperparameter-search budget as every other arm. Beating it means a precision-recall improvement whose bootstrap interval excludes zero at the 50-positive setting.

Raw-embedding hybrid. This is the strongest existing method and the direct analogue of Taubert et al., Reinders et al. and Pang et al.. It uses the same frozen encoder, the full raw embedding concatenated with clinical variables, and the same tree ensemble with the same tuning budget. It is the strongest competitor because it uses identical information and identical model capacity and differs only in the normative step. For this preclinical setting there is no reported result found on Lacuna for the raw-embedding hybrid, or for any baseline. All baseline numbers will be produced inside this study rather than quoted. Beating it means higher precision-recall area at 25 and 50 positives, plus a smaller drop under leave-one-site-out testing.

End-to-end fine-tuning. The encoder is fine-tuned on the labeled set with a small classification head, weight decay, early stopping, and clinical variables appended to the head. It gets the same compute budget and the same number of tuning trials.

Key ablations. These isolate the key ingredient:

  • Principal components of the raw embeddings, computed without covariate conditioning. This separates "low dimension" from "deviation from age-matched healthy peers".
  • The deviation scores fed to logistic regression instead of trees. This separates the representation from the classifier.
  • Normative models fitted on progressively smaller healthy pools. This measures how much unlabeled data the method actually needs.

Why not just…?

Why not just put a linear probe or penalized logistic regression on the raw embeddings plus clinical variables? This is a serious threat. Several hybrid gains reported in the literature may come from regularization rather than tree geometry. The proposal's claim is therefore deliberately about the normative transformation, not about trees. If logistic regression on deviation scores matches the tree ensemble, we report that, and the contribution becomes the representation. If a linear probe on raw embeddings matches the deviation pipeline, the core claim is falsified. The ablation grid above contains exactly these arms.

Why not just use clinical and laboratory variables and skip imaging? For some diseases this may well be the right answer. The clinical-only baseline is the first number reported, not a footnote. The decisive check is whether deviation scores add precision-recall area on top of clinical variables in held-out sites. A null result here is itself useful, because it says when preclinical imaging is not worth acquiring.

Why not just fine-tune the encoder with heavy regularization or data augmentation? Taubert et al. report that neural heads overfit badly under label scarcity where gradient-boosted heads do not, and Badjatiya et al. find hybrids beating end-to-end training. Those are different domains, though, so the fine-tuning arm is run here at matched budget. We expect fine-tuning to close the gap only at the largest positive count. If it wins at 25 to 50 positives, the premise is wrong.

Why not just fit the normative model and threshold the abnormality score, with no supervised classifier? This unsupervised arm costs nothing and will be reported. Its weakness is that it flags every kind of abnormality, including incidental and non-disease findings. The supervised stage is meant to learn which directions of deviation predict this particular outcome. The comparison measures how much that supervision is worth.

Decisive Checks

The idea succeeds if three things hold. At 25 and 50 positives, the deviation-score hybrid beats the raw-embedding hybrid on precision-recall area, with non-overlapping bootstrap intervals. It beats the clinical-only baseline under leave-one-site-out testing. The gap narrows as positives grow toward 200, which is the specific signature predicted by "normative modeling substitutes unlabeled data for labels".

The idea fails in a specific, publishable way if principal components without covariate conditioning do as well. That would mean only dimensionality mattered. It also fails if the gains vanish across sites, which would mean the normative model learned scanners rather than biology.