Short Screeners That Measure the Dimension, Not the Secret

Clinics increasingly replace long psychiatric questionnaires with a handful of items chosen by an algorithm. On the Healthy Brain Network cohort, three to five questions match or beat full clinician subscales. The usual goal of that choice is to pick the few measurements from which everything else can be predicted. That goal has a side effect. If five answers predict the whole battery, they also predict the answers the patient never gave, including the most sensitive ones: self-harm, substance use, psychotic experiences. The biometrics literature has already shown what happens when a compact, faithful representation is exposed through a similarity score. Attackers turn the score into a compass and rebuild the hidden record, and the more faithful the representation, the better the attack works.

We propose to choose screener items the way engineers choose sensor locations, with one change to the objective. The selected items should pin down a continuous clinical dimension, such as a mood or externalizing severity axis, while telling an outsider as little as possible about designated sensitive items beyond what that dimension already implies. Choosing this layer is deliberate. Any score, distance, or label a downstream system releases is computed from the selected answers, so the item set caps the worst-case leakage of every interface built on top of it. We expect clinical batteries to be redundant enough that many item subsets are nearly equally good at measuring the dimension but differ widely in what they give away. The first job of the project is to test that expectation, because if it fails, the rest of the approach fails with it.

Core Claim

Three bodies of work supply the pieces. Sensor selection picks $k$ of $n$ candidate measurements under a budget. It uses greedy maximization of diminishing-returns objectives with a $(1-1/e)$ guarantee for monotone functions, as in Near-optimal Observation Selection using Submodular Functions, or QR pivoting on a low-rank basis, as in PySensors. Both assume the consumer of the measurements is trusted. Dimensional psychiatry supplies the target. Linked dimensions of psychopathology and connectivity in functional brain networks recovers four continuous symptom dimensions from 111 symptoms and 3,410 brain connections in the Philadelphia Neurodevelopmental Cohort. That work motivates replacing categorical targets, but it has no budgeted selection. Accurate and efficient data-driven psychiatric assessment using machine learning does budgeted item selection, using recursive elimination followed by floating forward search. It targets categorical diagnoses and never asks what the chosen items reveal.

The distance-leakage literature supplies the threat and constrains the design. Similarity scores serve as fitness functions for genetic reconstruction in A Genetic Algorithm Enabled Similarity-Based Attack on Cancellable Biometrics, and longer, more accurate templates made that attack stronger. Continuous Euclidean and cosine distances admit linear-program inversion in Cancelable biometric schemes for Euclidean metric and Cosine metric. Distances to reference points triangulate the target in Secure binary embeddings of front-end factor analysis for privacy preserving speaker verification. Encryption does not help once the score is decrypted, per Template recovery attack on encrypted face recognition systems with unprotected decision using synthetic faces.

Together these results say that interface-level defenses leak whenever the interface is continuous and informative. The one thing an interface cannot exceed is the information in its input. Write $x_S$ for the answers to selected items $S$, $z$ for the true continuous dimension scores (defined from the full battery), and $p$ for the protected items. Any released output $g(x_S)$ satisfies $I(p; g(x_S)) \le I(p; x_S)$ by the data-processing inequality. The attacks above show this bound is nearly reached in practice. We therefore select

$$S^\star = \arg\max_{|S| \le k} \; I(z; x_S) \quad \text{subject to} \quad I(p; x_S \mid z) \le \tau,$$

where $I(z;x_S)$ measures how well the items determine the dimension and $I(p; x_S \mid z)$ is the excess leakage: what the items reveal about protected answers beyond what the dimension itself implies. The conditioning matters. If suicidal ideation genuinely loads on the mood axis, releasing a mood score reveals some of it no matter which items are used. Only the excess is under the designer's control.

The claim is that this constrained optimum reaches dimension fidelity close to the unconstrained optimum at substantially lower excess leakage. We also claim that simpler exclusion rules miss that frontier, because leakage is joint: individually innocuous items combine to triangulate a sensitive one, much as integer-programming attacks invert individually uninformative hash bits in Locality-Sensitive Hashing Does Not Guarantee Privacy!.

Research Program

The first component is estimating the two quantities. On the Healthy Brain Network item-level data we fit continuous dimensions on a training split. We use factor analysis over the full battery as the primary definition, and sparse canonical correlation as a secondary one following Xia et al. Protected items are a designated set of sensitive items, for example self-harm or substance-use questions where the battery contains them. Under a Gaussian proxy fitted to the item covariance, both mutual informations have closed forms in posterior covariances. That makes evaluating a candidate subset cheap.

Because the proxy may be wrong for ordinal answers, every selected subset is also audited with learned adversaries that the objective never sees. These are gradient-boosted predictors trained on a disjoint "public" split to predict each protected item from the released record, with and without access to the true dimension scores.

The second component is search. The difference between an informativeness term and a leakage term is not monotone, the case Krause et al. leave open. So we do not claim a greedy guarantee. We run two searches. The first follows the Konishcheva et al. filter step to shrink the pool to about 30 candidates, then enumerates every subset exactly; $\binom{30}{5}$ is about 142,000, so this is feasible. The second is a lazy greedy over the full pool with the constraint enforced at each step. Exact enumeration gives us the complete set of near-optimal subsets, which we call the "Rashomon set." That set is the object the core claim depends on. Enumeration also measures how far greedy falls from the true optimum.

The third component is the interface study. Downstream services rarely release raw answers. They release a continuous dimension score, a deviation from a normative model in the style of Breaking the Norm, distances to reference patient profiles, or a categorical label. For each interface we run the matching attack from the biometrics literature: triangulation from distances to $m$ reference profiles, linear-program inversion of distances, and genetic search using the score as fitness. We ask two questions. How much of the raw-record leakage does each interface recover? Does a low-leak item set stay low-leak under the strongest attack, as the data-processing argument predicts? This is also where categorical diagnosis enters as a lens. A categorical label is a crude privacy mechanism that throws away the continuum. We measure its price at the diagnostic boundary, where subthreshold cases overlap heavily with diagnosed ones.

The fourth component is replication in a second modality. In the Philadelphia cohort the "sensors" are brain connections. We select $k$ edges that estimate the Mood, Fear and Externalizing scores, with psychosis items as the protected set. This tests whether the frontier exists in the setting that fits the physical sensor-selection framing most literally.

Baselines

The naive reference points are the full battery (best fidelity, maximal leakage), random $k$-item subsets, and a categorical diagnosis release alone. Together they bracket the frontier.

The strongest existing selection method is the Konishcheva et al. elimination-plus-floating-search pipeline, run through its released modular code. We retarget it to maximize held-out correlation with the continuous dimension instead of diagnostic AUROC. It uses the same splits, the same candidate pool and the same tuning budget as our search. Only the objective differs.

For calibration, the file reports this pipeline's categorical performance on Healthy Brain Network. It reaches an AUROC of 0.865 for autism with 4 items, 0.840 for generalized anxiety with 5 items, and 0.810 for oppositional defiant disorder with 3 items. Our dimension scores, thresholded, should come within noise of these numbers on the same diagnoses, or the dimensional target is not clinically competitive. For fidelity to the continuous dimensions, the only reported figures are the Xia et al. brain–behavior correlations of r = 0.68–0.71. Those are canonical correlations, not short-form fidelity. There is no reported result found on Lacuna for dimension fidelity from budgeted item subsets, and none for protected-item leakage from psychiatric short forms. We therefore establish both baselines ourselves.

QR pivoting from PySensors, applied to the item loading basis, is the standard reconstruction-oriented selector. We expect it to be the leakiest, because it explicitly optimizes recovery of the whole battery.

Two simple privacy baselines receive the same enumeration budget. The first is correlation exclusion: drop any item whose marginal correlation with a protected item exceeds a threshold, then select normally. The second is select-then-noise: take the Konishcheva subset and add calibrated Gaussian noise to the released score.

Our ablations isolate the key ingredient in two ways. One replaces the joint, conditional leakage term with a sum of marginal per-item leakages. The other sets the leakage budget to infinity.

Beating a baseline means the following. At matched held-out dimension fidelity, within 0.02 in correlation, our subset must show lower excess leakage under the held-out adversary. We measure excess leakage as protected-item AUROC from the record minus AUROC from true dimension scores alone. The improvement must hold across cross-validation folds and across the protected items.

Why not just…?

Why not just add noise to whatever score gets released? Noise can buy the same leakage reduction only if the service is the sole channel. Stored short-form answers, audit logs and reused responses bypass it. Bassit et al. show that protection at one layer is undone once plaintext scores escape. Noise also degrades fidelity uniformly, and hardest at the diagnostic boundary. Still, this is a real competitor for the score-only interface. The select-then-noise baseline, compared at matched fidelity, decides it. If noise matches our frontier there, we will report that selection only matters when raw records are retained.

Why not just drop items that correlate with sensitive questions? This is the most serious threat, because it is cheap and clinicians would accept it. Its weakness is that it only sees one item at a time. Joint leakage from combinations, the effect behind the integer-programming inversions of Turati et al., is invisible to it. It also cannot tell leakage the dimension already implies from excess leakage, so it may throw away the best mood items for no privacy gain. The marginal-leakage ablation and the correlation-exclusion baseline test this directly. If either lies on our frontier, the joint objective is unnecessary, and we will say so.

Why not just release a categorical diagnosis? A one-bit output leaks less per query. Binary match interfaces still yield to boundary search, as noted in Beck and Beck's line of work. Categories also discard exactly the graded information that dimensional work, such as Xia et al. and RedDino, shows is clinically real. We quantify that cost as fidelity lost among near-threshold patients, set against the leakage saved.

Why not just use fewer items? Shrinking $k$ reduces leakage and fidelity together, along one curve. Our claim is that at a fixed $k$ there is a spread of subsets off that curve. Sweeping $k$ for the Konishcheva selector traces the curve, and our subsets must lie strictly below it in leakage.

Decisive Experiments

The first and most important check is the Rashomon spread on Healthy Brain Network. Among all enumerated 5-item subsets within 0.02 of the best held-out dimension fidelity, we measure how widely excess leakage varies under the held-out adversary. If the spread is narrow, the target dimensions and the protected items are too entangled for selection to help. In that case the proposal reduces to an interface-design study, which we would still report as a negative result. If the spread is wide, the second check asks whether the Gaussian-proxy objective finds the low-leak end without seeing the adversary. We measure the rank correlation between proxy leakage and adversary AUROC across subsets, and the gap between greedy and exhaustive search.

The third check is the interface study. For each release format and matching attack, we measure the fraction of raw-record leakage recovered. Using triangulation with a growing number $m$ of reference distances, we test whether continuous-distance interfaces approach the raw-record bound for leaky subsets but not for our low-leak subsets. That is the prediction that justifies placing the defense at selection.

The fourth check repeats the frontier analysis on Philadelphia cohort connectivity edges. It is judged by whether the same qualitative spread appears when the "sensors" are brain connections rather than questions. Throughout, the categorical AUROCs Konishcheva et al. report serve as a floor on clinical usefulness that no low-leak subset may fall below.