Many clinical prediction systems are built as pipelines. One stage picks a small set of brain regions or genes, a second stage turns those picks into derived measurements such as rates of change, and a third stage makes the diagnosis. The first stage usually uses group sparsity, a penalty that keeps or discards each feature as a whole unit across several related targets. Group sparsity reliably beats picking features target by target on small, high-dimensional datasets. The selection step is also a pipeline boundary, though, and it is the harshest one: a region that stage one zeroes out can never be used by the diagnostic stage, however useful it would have been there. This is the same "information starvation" that trace-back studies of language-processing pipelines document, moved into the feature-selection setting.
We propose to fix this by changing what the group is. Each feature's group will contain its weights in every stage that uses it, including the downstream weights on quantities derived from it, such as that region's slope and intercept over time. The question of whether a feature survives is then decided by all stages together. Two things are new here. The first is the grouping axis: features are grouped across sequential stages rather than across parallel tasks or the classes of a taxonomy. The second is a direct measurement of how much downstream error the upstream selection causes, which no prior work in the file reports.
Core Claim
Our hypothesis is that in pipelines where later stages consume transformed versions of selected features, much of the final-stage error comes from selection at stage one. We further hypothesize that a single group penalty spanning stage-one and stage-three weights recovers most of that loss. It should do so without giving up the compact, interpretable feature set that motivates pipelines in the first place, and it should beat the simpler fix of adding the final label as one more target in the stage-one regression.
The closest existing system is the three-stage mild cognitive impairment (MCI) pipeline of Predicting Future Clinical Changes of MCI Patients Using Longitudinal and Multimodal Biomarkers. It works in three steps. First, it selects regions of interest with an $\ell_{2,1}$-penalized regression across visits. Second, it fits polynomial trends to each selected region. Third, it feeds the results to a multi-kernel support vector machine. Information flows only forward, so stage-one mistakes cannot be corrected later. Relieving the Computational Bottleneck (BioMLN) attacks pipeline error from the other side. It passes support-vector-machine confidence margins as soft evidence into a Markov Logic Network, but its high-dimensional features are chosen in isolation from the downstream reasoning. Hierarchical Feature Selection with Recursive Regularization couples $\ell_{2,1}$ selections between parent and child nodes of a class tree. It does not handle a structure in which one stage's outputs become the next stage's inputs. Our proposal sits in the gap all three leave: joint selection along the processing direction of a pipeline.
Research Program
Let $X$ be the matrix of subjects by $d$ region-level measurements, and let $Y$ be cognitive scores at several visits, which is the stage-one target. Let $\Phi(X)$ be the stage-two derived features: for each region $i$, a small block of trend coefficients across modalities. Let $y$ be the final conversion label. We fit stage-one weights $W^{(1)}$ and stage-three weights $W^{(3)}$ jointly:
$$\min_{W^{(1)},W^{(3)}} \; \mathcal{L}_1(XW^{(1)},Y) + \lambda\,\mathcal{L}_3(\Phi(X)W^{(3)},y) + \gamma\sum_{i=1}^{d}\sqrt{\|w^{(1)}_i\|^2 + \alpha\,\|w^{(3)}_i\|^2}$$
The symbols are as follows. Here $w^{(1)}_i$ is region $i$'s row of stage-one weights across visits, and $w^{(3)}_i$ is the block of stage-three weights on region $i$'s derived trend features. The parameter $\lambda$ trades off the two losses, $\alpha$ rescales the downstream block so the two stages contribute comparably to the group norm, and $\gamma$ sets overall sparsity.
Because stage two is a fixed least-squares projection, the third stage is linear in $\Phi(X)$ and the whole objective stays convex. It can be solved with the same reweighting scheme as Efficient and Robust Feature Selection via Joint ℓ2,1-Norms Minimization (RFS). Region counts here are around a hundred, so that paper's matrix-inversion bottleneck does not bind. In a second variant, we replace the convex group norm with the capped penalty of Multi-Stage Multi-Task Feature Learning (MSMTFL). That paper shows the capped penalty reduces the estimation bias of convex relaxations. It also lets features that are strongly supported by either stage escape shrinkage.
Alongside the method, we build a diagnostic for selection-induced error. For a fitted pipeline, we re-add each discarded region to stage three alone and refit. We then record the downstream gain on held-out folds and count "starved" features, meaning those whose re-addition helps by more than the fold-to-fold noise. We summarize this as the fraction of the gap between the pipeline and a no-selection stage three that is closed by re-adding starved features. This brings the false-negative trace-back of When it's all piling up to the feature level. It also gives a pipeline-level counterpart to the propagation metrics that ReCAPA defines for agents.
As a secondary test of generality, we will apply the same grouping in an event-extraction pipeline of the kind BioMLN uses. There, trigger and argument classifiers share lexical and syntactic features. We group each feature's trigger-stage and argument-stage weights and keep BioMLN's downstream joint inference unchanged.
Baselines
The naive reference is the feed-forward pipeline exactly as in the MCI paper. It runs $\ell_{2,1}$ selection on cognitive scores, then trend fitting on the selected regions, then a classifier. That paper reports 78.4% conversion accuracy on 88 MCI subjects. We will re-run it on our splits rather than reuse that number, because the cohort is tiny and our folds differ. A second reference is stage three trained on all regions with no selection. It bounds what selection can cost, and it shows whether selection helps at all under this sample size.
The strongest existing selector is RFS's joint $\ell_{2,1}$ regression. The file reports it beating ReliefF, F-score, t-test, information gain and mRMR by 5% to 12% accuracy at 20 features on six bioinformatics datasets. We will also run MSMTFL, which reports an MRI regression normalized mean squared error of 0.5539 against 0.6651 for Lasso. We will also run tgLASSO from Temporally-Constrained Group Sparse Learning for Longitudinal Data Analysis, which reports mean correlation 0.613 for MMSE (a standard cognitive-screening score) against 0.587 for group Lasso. Each of these serves as the stage-one selector inside the same pipeline.
The ablation that isolates our key ingredient is the "extra column" model. It appends the conversion label as one more target in the stage-one $\ell_{2,1}$ regression on raw regions. This gives downstream supervision a voice in selection without the cross-stage grouping over derived features.
All methods get the same folds, the same nested cross-validation budget over penalty weights (a fixed grid of equal size per method), and the same stage-three model family. When the comparison calls for it, we also report the original multi-kernel SVM on top of each selected set. To count as beating the baselines, our method must show higher conversion accuracy, or higher cognitive-score correlation, than both the feed-forward pipeline and the extra-column model, outside paired-fold variability. It must do this at an equal number of selected regions. It must also show a lower starved-feature fraction.
Why not just…?
Why not just add the final label as another task in the stage-one regression? This is the real threat, and it is why the extra-column model is a primary ablation rather than an afterthought. It differs from our proposal only when stage three's useful signal lives in derived quantities, such as the rate of decline, rather than in raw cross-sectional values. If extra-column matches us, then on these data cross-stage grouping adds nothing beyond downstream supervision, and we will report that. We expect a difference specifically for regions whose levels are uninformative but whose slopes predict conversion. The starved-feature diagnostic will show whether such regions exist.
Why not just select features independently at each stage, or loosen the stage-one threshold to keep more regions? RFS and MSMTFL already show that independent selection loses to joint selection on small samples. A looser threshold trades starvation for overfitting in the 88-to-445-subject regime. We will sweep the number of retained regions for the feed-forward pipeline and plot accuracy against size. Our claim fails if some looser threshold matches our accuracy at comparable size.
Why not drop selection and train one end-to-end model? The no-selection stage-three baseline tests exactly this, using a linear model and, as a check, a small neural network. On the sequence-labeling side, Biomedical Event Extraction as Sequence Labeling shows that collapsing a pipeline can win but causes under-prediction and loses cross-sentence events. More fundamentally, it gives up the named region list that clinical use requires. If the unselected model wins clearly, selection is the wrong goal for this cohort, and we would say so.
Proposed Experiments
The primary testbed is ADNI (the Alzheimer's Disease Neuroimaging Initiative), using two cohorts. The first is the 445-subject cohort of 93 grey-matter regions with visits at baseline, 6, 12 and 24 months, as used by tgLASSO. The second is the multimodal MCI conversion setting with MRI, PET and cognitive scores. Outcomes are 24-month MMSE and ADAS-Cog correlation (the cited papers report 0.786 and 0.777 for the MCI pipeline, and 0.613 MMSE mean correlation for tgLASSO, on different cohorts, so they are not directly comparable), along with conversion accuracy, sensitivity and specificity. We will also report selection stability as the bootstrap Jaccard overlap of selected regions, together with the starved-feature fraction.
The secondary testbed is BioNLP'13 event extraction, where BioMLN reports F1 53.61 and a 10-point gain on nested regulation events over its SVM pipeline. The question there is whether trigger–argument grouped selection adds F1 at the same feature budget, holding the downstream logic network fixed.
The decisive check has three parts, all of which must hold. First, the starved-feature diagnostic must show that feed-forward selection discards regions that measurably help stage three. Second, our method must recover a substantial share of that gap. Third, it must beat the extra-column ablation at equal selection size. If the first part fails, selection-induced propagation is not a real problem in these pipelines, and that negative result is itself worth publishing. If the first part holds but the third fails, the simpler fix suffices and becomes our recommendation.
References
- Predicting Future Clinical Changes of MCI Patients Using Longitudinal and Multimodal Biomarkers
- Relieving the Computational Bottleneck: Joint Inference for Event Extraction with High-Dimensional Features
- Hierarchical Feature Selection with Recursive Regularization
- Efficient and Robust Feature Selection via Joint ℓ2, 1-Norms Minimization
- Multi-Stage Multi-Task Feature Learning
- When it's all piling up: investigating error propagation in an NLP pipeline
- ReCAPA: Hierarchical Predictive Correction to Mitigate Cascading Failures
- Temporally-Constrained Group Sparse Learning for Longitudinal Data Analysis
- Biomedical Event Extraction as Sequence Labeling