Subtype-Balanced Topological Codebooks for Finding Rare Disease Subtypes

Diseases like Alzheimer's and frontotemporal dementia do not follow one path. Different patients lose different brain structures in different orders. Methods that recover these hidden subtypes from scans currently describe each brain with a few dozen regional volumes. Shape changes that volume misses, such as a ventricle wall thinning in one place or the gaps between cortical folds widening, therefore go unseen. Persistent homology can measure exactly this kind of shape change. It is a tool that records when holes, loops, and gaps appear and vanish as you sweep a threshold across an image, and it outputs a variable-size set of (birth, death) points called a persistence diagram. To feed these diagrams to standard models, people usually compress them with a learned codebook: a small set of reference points in the diagram plane, with each diagram summarized by how much of its mass falls near each reference point.

The problem we target is that the codebook is learned from the pooled population. It spends its reference points where most patients' features lie, so it underrepresents the shape signatures of a subtype that makes up a small fraction of the cohort. In practice that fraction is often small. One longitudinal Alzheimer's analysis found an aggressive subtype at 14% prevalence against 86% for the slow one. We propose to learn the codebook jointly with the subtype model, and to refit it so that every inferred subtype gets an equal share of reference points. The rare subtype's geometry then survives compression. Each reference point stays tied to a named anatomical structure and a physical length scale, so the subtypes the model finds remain readable as ordered anatomical events.

Core Claim

Codebooks fit to the pooled distribution of diagrams are biased against minority subtypes. That bias, and not a lack of topological signal, is what would keep shape-based features from improving subtype discovery in imbalanced cohorts. Alternating between fitting the subtype model and refitting the codebook on a subtype-balanced mixture of diagrams should recover minority subtypes that both volume-based and pooled-codebook pipelines miss, without degrading majority-subtype staging. The new element is the feedback from inferred, unlabeled subtype structure into the vectorization step. Neither balancing a codebook by known class labels nor adding topological features to a progression model would supply that feedback.

Prior Work and Where This Sits

The subtype model we build on is Subtype and Stage Inference (SuStaIn), introduced in Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference. It models each subtype as an ordered sequence of biomarkers crossing z-score thresholds relative to healthy controls, and fits the subtypes from cross-sectional data by sampling. Its extensions handle missing modalities (Multimodal subtypes identified in Alzheimer's Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference) and physical timescales (Subtype and Stage Inference with Timescales). All three operate on scalar regional summaries. The timescale paper reports the 86%/14% subtype split that motivates our imbalance concern, but it has no mechanism that protects small subtypes. We keep SuStaIn's inference unchanged and replace its inputs.

The vectorization method comes from ATOL: Measure Vectorization for Automatic Topologically-Oriented Learning. ATOL runs a Lloyd-style quantization on the mean measure of training diagrams and sums localized contrast functions to produce fixed-length vectors. Its authors note that it tends to miss discriminative features at the edges of the measure. That is the failure we expect to worsen under subtype imbalance. Related codebook work (Persistence Bag-of-Words for Topological Data Analysis; Persistence codebooks for topological data analysis) shares the same pooled-density objective. We keep ATOL's contrast functions and change only the measure the centers are fit to.

The filtration that makes codewords interpretable comes from A Topological Gaussian Mixture Model for Bone Marrow Morphology in Leukaemia. Its signed-distance persistent homology assigns negative values inside a structure and positive values outside. Negative-birth features then read as wall or tissue thickness, and positive-birth features read as gap or void width. With this filtration the authors reached 88.9% leave-one-out accuracy against roughly 66.7% for classical morphometry. That result motivates the claim that shape topology carries stage information that volume averages lose. It was obtained on 27 mouse samples with fixed stage labels and no subtyping, so it is weak evidence for human neuroimaging, and we treat it that way.

Two papers constrain the design. Integrating Clinically-Relevant Features into Skin Lesion Classification found that forcing a model to predict clinician-defined concepts hurt accuracy on intermediate classes. We therefore do not impose hand-named concepts. Interpretability comes from the filtration's geometry: each codeword is a region of (birth, persistence) space for a specific structure. Learned end-to-end topological layers (Deep Learning with Topological Signatures; Integrating Multi-scale and Multi-filtration Topological Features for Medical Image Classification) improve static classification. They optimize for labels and lose the per-feature anatomical reading that a staging model needs, so we use them only as a comparison.

Research Program

For each subject, we segment the lateral ventricles, hippocampi, and cortical cerebrospinal fluid from T1 MRI. We then compute signed-distance persistence diagrams in dimensions 0, 1 and 2 for each structure.

An initial ATOL codebook of $m$ centers per structure is fit on the pooled diagrams of patients. Each codeword's activation is z-scored against healthy controls and passed to SuStaIn as a biomarker. SuStaIn returns, for each patient $i$, a posterior probability $w_{ik}$ of belonging to subtype $k$.

We then refit the centers on a balanced measure in which each of the $K$ subtypes contributes equally, regardless of its size:

$$ \bar\mu_{\text{bal}} = \frac{1}{K}\sum_{k=1}^{K} \frac{\sum_i w_{ik}\,\mu_i}{\sum_i w_{ik}}, $$

where $\mu_i$ is patient $i$'s persistence diagram treated as a point measure. We optionally subtract the control mean measure, so that centers go where patients differ from healthy subjects rather than where all brains look alike. The new codewords define new biomarkers. SuStaIn is refit, warm-started from its previous sequences. We alternate these two steps until the held-out likelihood stops improving. The number of subtypes is chosen by cross-validated likelihood, as in the original SuStaIn.

To keep the loop from inventing subtypes that only it can see, we cap the number of iterations. We also report results from every iteration, so any drift is visible.

Interpretation works as follows. Each codeword is a small region of (birth, persistence) space for a known structure, so a SuStaIn event reads as a sentence of the form "ventricular gaps of radius 3–5 mm become abnormal at stage 4 in subtype 2." We verify that reading by mapping the representative cycles of a codeword's contributing points back onto the subject's image. Two raters then check whether the highlighted locations match the stated structure and scale.

Baselines

The naive reference point is volume-only SuStaIn, which runs on standard regional volumes from the same segmentations. This is what the published pipelines do. Its strongest reported number in the file is "95% accuracy (for GRN vs. MAPT) based solely on brain atrophy patterns" on GENFI, from Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference. That headline comes from the natural cohort composition. The file reports no result under controlled subsampling of the minority mutation group. Our proposal counts as beating this baseline only if it matches it on the natural split and exceeds it once the minority group is shrunk.

The strongest existing topological competitor is pooled-ATOL SuStaIn: the same diagrams, the same number of codewords, and an ATOL codebook fit once on the pooled measure. This isolates the key ingredient, because the only difference from our method is the balanced, iterated refit. No reported result on Lacuna exists for ATOL features inside a progression model. ATOL's reported results are graph benchmarks, for example 67.1% on REDDIT-5K.

Two further ablations separate our mechanism from simpler fixes. In the label-balanced codebook, centers are fit on a measure balanced by observed diagnosis (control, mild impairment, dementia) rather than by inferred subtype. In the inverse-density codebook, centers are fit with weights that up-weight low-density regions of the pooled measure, and no subtype feedback is used.

The fourth comparison is a supervised black box: an end-to-end differentiable topological layer in the style of Deep Learning with Topological Signatures, trained to predict conversion from mild cognitive impairment to dementia. It is compared to our model only on conversion prediction. The file's reference number for this task is the multimodal recurrent model in Predicting Alzheimer's disease progression using multi-modal deep learning approach, at "81% accuracy (AUC = 0.86)". That model uses longitudinal multimodal inputs, so we treat it as context rather than a matched target.

Fairness rules apply to every arm. All arms use identical subjects and folds. Feature count is matched, with the total number of codewords equal to the number of regional volumes. All arms get the same sampling budget and the same procedure for choosing the number of subtypes. Codebook size is tuned on the same grid for all topological arms.

Why not just…?

Why not just use regional volumes, as SuStaIn already does? This is the main threat. The file's evidence that topology beats morphometry comes from a 27-sample mouse study and from static image classification, not from brain progression. If volume-only SuStaIn matches us on minority-subtype recovery in the controlled GENFI subsampling experiment, the topological premise fails for this application, and we will report that outcome.

Why not just balance the codebook by diagnosis label, or oversample minority patients? Subtypes are latent. In ADNI, both the 86% and 14% subtypes contain patients at every diagnosis. Balancing by diagnosis therefore equalizes stages, not subtypes. The label-balanced ablation tests this directly. If it recovers the rare GENFI group as well as our method does, the subtype feedback is unnecessary.

Why not just put more codebook centers in low-density regions, with no feedback loop? This alternative might work as well as ours. Rare-subtype features probably sit in sparse regions of the pooled measure, and inverse-density weighting is cheaper. Its weakness is that it also amplifies scanner noise and segmentation artifacts, which sit in sparse regions too. The inverse-density ablation measures minority recall together with control-group false-abnormality rates. If it ties our method on both, we will recommend it instead and drop the joint fit.

Why not just train an end-to-end topological network and explain it with saliency maps? Such a network can predict conversion, but it produces no ordered event sequence and no per-subtype stage. The breast-ultrasound work notes that saliency maps are qualitative and noisy (Lesion-Centric Latent Phenotypes from Segmentation Encoders for Breast Ultrasound Interpretability). We report the network's conversion area under the curve as an upper reference. If our method stays within a few points of it while also yielding stages, the trade is worthwhile.

Proposed Experiments

Synthetic test. We modeled this test on the 500-subject synthetic progression benchmark of Multimodal subtypes identified in Alzheimer's Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference. We generate 3D phantoms with two subtypes. One subtype expresses a gap-widening event that changes topology but leaves volume constant. We sweep its prevalence from 50% down to 5%. The metrics are minority-subtype recall, subtype adjusted Rand index, and Kendall correlation of recovered event order with the ground truth. Pooled ATOL should lose the minority subtype at a higher prevalence than our method does. The measured gap between these two breakdown points is the effect size of the whole proposal.

Controlled-imbalance test. In GENFI, mutation group is a genetic ground truth for subtype. We subsample MAPT carriers to 30%, 15% and 7% of the mutation-carrier sample, across twenty random draws, and compare all arms on minority recall and balanced accuracy. A positive result requires our method to hold minority recall as the group shrinks while volume-only and pooled-ATOL SuStaIn decline.

ADNI test. We report whether a fast, small subtype like the one in Subtype and Stage Inference with Timescales appears, with what prevalence, and how stable it is across bootstrap fits. We then report whether subtype-by-stage assignment predicts conversion beyond stage alone, using a likelihood comparison in the spirit of SuStaIn's original model-fit tests.

Interpretability check. Raters score whether the image locations behind each codeword match its stated structure and scale. They also score whether the event sequences are anatomically plausible, blind to the arm that produced them.