Biomarkers of Alzheimer's disease switch on at different times. Amyloid in spinal fluid rises early and then flattens out. Brain metabolism falls in the middle stages. Brain shrinkage speeds up late. Cognitive scores decline last. Most methods that pick a small set of predictive biomarkers from many candidates fit linear models, or models that are linearized around one point. A linear model treats a biomarker that has flattened out as carrying no information, so it tends to throw that biomarker away. We propose to track each patient's hidden disease stage over time with a nonlinear filter. The filter is paired with a penalty that removes whole groups of biomarkers, and both are fit together in one objective. The key design choice is how uncertainty is pushed through the S-shaped curves that link stage to biomarker values. We use the unscented transform, which pushes a few carefully placed sample points through the true curve instead of a tangent line. This lets the selection step see that a flattened biomarker is still informative over the part of the stage range where the patient might be.
The concrete claim can be stated in two sentences. When biomarkers follow S-shaped curves over a hidden disease stage, feature selection driven by a linearized (Jacobian-based) filter will prune biomarkers whose curves have already flattened. An unscented filter keeps them, and it also handles missed scans and skipped lumbar punctures without dropping patients. We test this on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, using the full cohort with its missing data rather than the 88-patient complete-case subset used in earlier longitudinal work.
Core Claim and Model
Each patient $p$ has a low-dimensional hidden state $x_t = (s_t, r)$, where $s_t$ is disease stage at visit time $t$ and $r$ is a log progression rate. Stage moves forward monotonically: $s_{t+\Delta} = s_t + \Delta \, e^{r}$ plus small process noise. Each biomarker $k$ belongs to one modality: regional MRI volume, regional FDG-PET uptake, CSF amyloid or tau, APOE-adjusted baselines, or cognitive subscores. Its value is generated as
$$y_{k,t} = a_k + b_k \,\sigma\!\left(\frac{s_t - c_k}{w_k}\right) + \epsilon_{k,t},$$
where $\sigma$ is the logistic function, $a_k$ is the biomarker's healthy level, $b_k$ is its total disease-related change, $c_k$ is the stage at which it changes fastest, $w_k$ is how gradual the change is, and $\epsilon_{k,t}$ is Gaussian measurement noise. If a biomarker is missing at a visit, it simply contributes nothing to that visit's update.
The parameters are learned by maximizing the filter's marginal likelihood minus a group penalty:
$$\max_{\theta} \; \sum_p \log \hat p_{\mathrm{UT}}(y^{(p)}_{1:T} \mid \theta) \; - \; \lambda \sum_{g} \| b_g \|_2 .$$
Here $\hat p_{\mathrm{UT}}$ is the likelihood the unscented Kalman filter computes from its prediction errors at each visit. $b_g$ collects the amplitudes $b_k$ of all biomarkers in group $g$, and $\lambda$ sets how many groups survive. A group is one anatomical region measured in both MRI and PET, or one fluid or cognitive assay. The penalty therefore keeps or discards a region across both imaging modalities together, much as Zhang and Shen selected regions that were consistent across MRI and PET. A group whose $\|b_g\|_2$ is driven to zero is deselected.
Two features make this design work rather than decorate it. First, the state has only two or three dimensions, so the unscented transform needs just five to seven sigma points per visit. This sidesteps the $2n+1$ cost problem that limits the Unscented Autoencoder in high-dimensional latent spaces. Second, the model changes how rigid joint sparsity behaves. The penalty still selects biomarkers all-or-nothing. But how much a selected biomarker matters at a given visit is set by the slope of its curve over the patient's uncertain stage, so its relevance varies with stage without any extra per-visit penalty.
Why the Unscented Transform Is the Ingredient, Not a Garnish
Take a patient whose estimated stage sits on an amyloid plateau. A linearized filter (the extended Kalman filter) evaluates the curve's slope at the mean stage, finds it near zero, and concludes that the biomarker tells us nothing. That near-zero value flows into the likelihood gradient for $b_k$, so the group penalty sees little reason to keep the group. This is the same failure as the $y=x^2$ example in Dual Estimation and the Unscented Transformation, which introduces sigma-point propagation and shows that linearizing at the mean misses curvature. The unscented transform places points on both sides of the mean stage. If the stage posterior overlaps the steep part of the curve, the biomarker's informativeness shows up. We reuse this propagation directly. We also borrow from Extended and Unscented Gaussian Processes, which embeds sigma-point linearization inside a learned likelihood objective and reports that this objective can diverge for sharply non-smooth likelihoods. Our logistic curves are smooth, but we still add damped updates as that paper advises. Deterministic sigma points also give a noise-free objective, which suits the proximal-gradient steps that group penalties need. Unscented Autoencoder motivates this: it replaced random sampling with sigma points to reduce gradient noise and reported CelebA FID improving from 65.86 to 38.48.
The mechanism also yields a scientific prediction that could be wrong. Ye et al. applied L1-penalized logistic regression with stability selection to baseline ADNI data. They found that adding CSF tau and amyloid gave no significant AUC gain over their 15-feature MRI, cognitive, and genetic signature. Our hypothesis is that part of this null result comes from applying a linear model to biomarkers that are already saturated in an MCI cohort. If the unscented model also discards CSF groups, the hypothesis is wrong, and we will report that.
Relation to Prior Work
Two ADNI lines of work motivate this proposal because they expose the gap.
Temporally-Constrained Group Sparse Learning (tgLASSO) selects features jointly across visits. It uses a group penalty plus smoothness penalties on how weights change between visits. It remains a linear map from features to scores, and each feature is selected for all visits or none.
Predicting Future Clinical Changes of MCI Patients chains three stages: group-sparse region selection, then polynomial trend fitting, then a multi-kernel SVM. Nothing learned in the later stages feeds back into selection, and the method needs complete data, which leaves 88 subjects. Our model replaces all three stages with one filter-plus-penalty objective and uses every available visit.
SuStaIn constrains our design as well as motivating it. It shows that ADNI contains subtypes with different biomarker orderings, fit with a linear z-score model and MCMC on cross-sectional data. A single cascade shared by all patients is therefore misspecified. We address this in one arm of the program by allowing a mixture of two or three orderings, with a separate filter for each ordering. SuStaIn does not select features, does not filter longitudinally, and does not map stage to calendar time. Our progression rate $r$ does that last mapping directly.
Stabl adds permuted noise features to calibrate false-discovery control in sparse multi-omic selection. We borrow its noise-injection idea to estimate how often our method selects null groups. We do not borrow its linear models.
Research Program
The work has three parts.
The first part is a simulator. It generates a cascade of 100–300 biomarkers with known curve parameters, irregular visits, and realistic missingness in each modality. Some truly relevant biomarkers are placed so that they saturate early in the simulated cohort. This gives ground truth for selection that ADNI cannot.
The second part is the estimator. It combines a square-root unscented Kalman filter over $(s, r)$ with automatic differentiation through the filter recursion, proximal group-soft-thresholding steps on $b$, and a path over $\lambda$. The mixture-of-orderings variant is included here.
The third part is the ADNI study. We use the full longitudinal cohort with MRI regional volumes, FDG-PET regional uptake, CSF measures, and cognitive subscores, keeping all participants with at least two visits of any modality. We forecast 24-month MMSE and ADAS-Cog by running the filter forward and passing the predicted stage through the learned curves for those scores. We also predict MCI-to-AD conversion. Selection stability is measured by bootstrap selection frequencies, and false selection by Stabl-style permuted decoy groups.
Baselines
The naive reference point is carrying the last observed cognitive score forward to 24 months, plus a linear regression on baseline MRI only. Ye et al. report an MRI-only AUC of 0.7877 for conversion, which anchors how much any multimodal method should add.
The strongest existing selection method is tgLASSO. It is the most direct prior method for joint feature selection across visits on ADNI, and it reports mean correlations of 0.613 for MMSE and 0.639 for ADAS-Cog across visits, against 0.569 for plain Lasso on MMSE (tgLASSO). Because it needs complete data, we run it in two ways: on a complete-case subset, and on the full cohort after multiple imputation. The strongest reported end-to-end longitudinal multimodal result is Zhang and Shen's pipeline. It reports 24-month correlations of 0.786 for MMSE and 0.777 for ADAS-Cog, and 78.4% conversion accuracy, on 88 complete-case subjects (Zhang and Shen). For conversion from baseline, the strongest static result is Ye et al.'s AUC of 0.8587. Our splits will differ from theirs, so those numbers serve only as orientation. We will reimplement both pipelines and run them on our own splits, with the same cross-validation folds, the same regional features, and the same number of tuning configurations (about 50 settings per method, chosen on inner folds).
Beating a baseline means one of two things. Either we win on held-out 24-month correlation in the complete-case subset, where the comparison is fair. Or we lose nothing there and win clearly on the full cohort, where the baselines must impute. In both cases, selection stability must not be worse.
Three ablations isolate the proposal's key ingredient. The first is the same model and penalty fit with an extended Kalman filter instead of the unscented one; this is the central comparison. The second is the same model fit by Monte Carlo filtering, with 10, 50, and 500 particles, at matched wall-clock time. The third replaces the logistic curves with linear curves; in that model the unscented and extended filters give identical results, so any difference in the full model must come from curvature.
Why not just…?
Why not just impute missing modalities and run tgLASSO or Stabl? This is the most serious threat to the prediction claim. On forecasting accuracy alone, imputation plus a well-tuned linear group penalty may well match us. Our claim therefore rests mainly on selection. Linear models should systematically miss relevant biomarkers that saturate early, and we can measure this directly in the simulator, where ground truth is known. On ADNI, the matching check is whether CSF and early-change groups survive in our model but not in the imputed tgLASSO. If imputed tgLASSO matches us on both prediction and simulator support recovery, we will narrow the contribution to handling missing data in one principled model.
Why not just use an extended Kalman filter, or Monte Carlo filtering, with the same nonlinear model? The extended filter is exactly the ablation our mechanism predicts will fail, because on plateaus its slopes are near zero. The simulator directly measures how often it prunes early-saturating groups. Monte Carlo is a real competitor. With a two- or three-dimensional state, a few hundred particles are cheap and more accurate than sigma points. Its weakness is noise in the objective and its gradients, which makes proximal group-penalty steps and the $\lambda$ path jittery. We will test this by comparing selection-path stability and final support at matched compute. If 50 particles give equally stable supports, we will report that the unscented transform's advantage is convenience rather than accuracy.
Why not just fit SuStaIn and then select features on its stage estimates? SuStaIn already models asynchronous cascades and subtypes. Running it and then applying group-penalized selection to stage-conditioned residuals is a sensible two-step alternative. It has three weaknesses: it is cross-sectional, it gives stages without calendar-time rates, and its z-score model is linear in stage, so the saturation problem appears again. We include this two-step pipeline as a further comparison. The decisive measure is 24-month forecast error, which the two-step method can only produce through a separate trend model.
Decisive Checks
The simulator gives the cleanest test. We will plot selection recall for early-saturating relevant groups against where they saturate, for the unscented filter, the extended filter, Monte Carlo filtering, and imputed tgLASSO. The claim holds only if the unscented curve stays high while the extended and linear curves fall, and the decoy-based false-selection rate stays below 10%.
On ADNI, three results together decide the proposal. First, a non-inferior 24-month MMSE and ADAS-Cog correlation against reimplemented tgLASSO and the Zhang–Shen pipeline on the complete-case subset. Second, a significant gain on the full cohort, supported by the larger number of patients we can use. Third, bootstrap selection frequencies showing which fluid and early-change groups the unscented model keeps and the extended-filter ablation drops.
A secondary check is whether adding CSF measures improves conversion AUC under our model when it did not under a linear model on the same splits. A positive result would suggest that Ye et al.'s null finding reflects how the model was specified rather than the biology. A negative result would be reported as evidence against our saturation hypothesis.
References
- Predicting Future Clinical Changes of MCI Patients Using Longitudinal and Multimodal Biomarkers
- Unscented Autoencoder
- Dual Estimation and the Unscented Transformation
- Extended and Unscented Gaussian Processes
- Sparse learning and stability selection for predicting MCI to AD conversion using baseline ADNI data
- Temporally-Constrained Group Sparse Learning for Longitudinal Data Analysis
- Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference
- Stabl: sparse and reliable biomarker discovery in predictive modeling of high-dimensional omic data