When a clinical decision-support service sits outside the hospital, such as a cloud diagnostic model, a referral partner, or a payer's triage tool, it usually receives either the whole patient record or a fixed redacted subset. Neither choice matches how diagnosis works. A clinician does not need every attribute to act. They need enough to tell apart the diagnoses that would lead to different treatments. This proposal treats each patient attribute the way reliability engineers treat a sensor. A disclosure policy decides which "sensors" the outside model can read, and a poor policy makes different diseases look identical to it, just as a sparse sensor layout makes different mechanical faults produce the same readings. The system we propose releases attributes to the outside model one at a time. At each step it picks the attribute that best separates the diagnoses still in contention, relative to how much that attribute adds to the patient's re-identification risk. It stops as soon as every remaining candidate diagnosis calls for the same action.
The opportunity comes from two literatures that have not met. Adaptive test selection already shows that fewer, well-chosen observations can beat full panels: Timely Clinical Diagnosis through Active Test Selection reports roughly 45% accuracy with all features versus 68% with active selection on Pima diabetes. That work assumes full plaintext access and uses binary tasks. Privacy architectures such as E-Tenon already enforce attribute-level, token-gated access, but they treat each request as a static, one-time retrieval and never ask what the withheld fields do to diagnosis. Our system connects the two. It produces a patient-specific sequence of disclosure requests that an access-control layer like E-Tenon could enforce, and it bounds the ambiguity those withholdings leave behind.
Core Claim
Privacy-driven withholding creates artificial ambiguity zones: groups of diagnoses that become indistinguishable only because of what was hidden. These zones can be measured from data. Much of the ambiguity they create does not matter for the decision, because the confused diagnoses share a treatment. A disclosure policy that resolves only the decision-relevant ambiguity, and that prices disclosures by their joint re-identification risk rather than a per-attribute cost, should reach the same action accuracy as full disclosure at substantially lower re-identification risk. It should also beat both static minimal-feature subsets and active selection with an additive privacy cost.
The fault-isolation literature supplies the lens. Data-Driven Fault Isolation in LTI Systems shows that when a fault's signal is cancelled internally before it reaches the sensors, the observed residual lands on the intersection of several fault hypotheses, so the faults cannot be told apart passively. The engineering response is to report the whole equivalence set rather than guess. Fault Separation Based on an Excitation Operator breaks such ties by active probing, but only by giving up something, namely yaw control. We carry both ideas over. Withheld attributes play the role of the cancelled signals, and each new disclosure is an active probe whose cost is privacy rather than control authority. Neither paper handles probabilistic, nonlinear clinical data, so what we borrow is the structure of the problem, not the geometric method.
Research Program
Let $S$ be the set of attributes disclosed so far and $x_S$ their values. A calibrated classifier trained on masked inputs gives a posterior $p(d \mid x_S)$ over diagnoses $d$. The ambiguity set $A(S)$ contains every diagnosis whose posterior exceeds a threshold $\tau$.
Each diagnosis maps to a clinical action $a(d)$, taken from guideline-level groupings. The protocol stops when all members of $A(S)$ share one action, or when the privacy budget runs out. In the second case it returns $A(S)$ as an explicit equivalence set for the clinician, mirroring engineering ambiguity zones.
Re-identification risk $R(S)$ is the fraction of a reference population whose disclosed tuple $x_S$ is shared by fewer than $k$ records. This quantity is deliberately non-additive, because two harmless attributes can become identifying together.
The next attribute is chosen by
$$j^\ast = \arg\max_{j \notin S} \frac{\mathbb{E}_{x_j}\big[H_a(S) - H_a(S \cup \{j\})\big]}{R(S \cup \{j\}) - R(S) + \epsilon},$$
where $H_a(S)$ is the entropy of the induced distribution over actions, not diagnoses, and $\epsilon$ keeps the ratio finite. We use action entropy because separating two diagnoses with the same treatment is wasted disclosure. The expectation over the unseen value $x_j$ is computed with a conditional generative model fit on training data. This matches ACTMED's predict-then-score design, but the predictive model is trained rather than an LLM surrogate, so results can be reproduced.
The analysis part of the program produces a disclosure-induced confusion map for each dataset. For every attribute subset reachable under a privacy budget, we record which diagnosis pairs merge. From this we report how much ambiguity is artificial, meaning present under the policy but not under full disclosure, and how much of that artificial ambiguity is action-irrelevant. This map is the clinical counterpart of listing transmission-zero blind spots.
Cryptographic enforcement is treated as an existing layer. The protocol emits one attribute-token request per step, and we count requests and round trips, but we do not build new cryptography. The single-shot encrypted classifier of Lightweight Multi-Class SVM Diagnosis with Privacy Preservation evaluates a linear model on a complete 34-feature encrypted vector. We ask instead which of those 34 features need to leave the hospital at all, which is a different kind of protection. It could be combined with ciphertext inference, since fewer released features also means a smaller encrypted vector to evaluate.
Baselines
The naive reference is full disclosure: the same classifier sees every attribute. It sets the ceiling on action accuracy and the worst case for risk $R$. On the six-class dermatology data, the file reports 97% accuracy for a one-versus-all linear SVM in both plaintext and ciphertext (Sensors 2023). We rerun that SVM, plus our calibrated classifier, on identical splits so that ceiling applies directly.
The strongest existing competitor is ACTMED with an additive privacy cost. We take the information-gain test selection and KL stopping rule of Timely Clinical Diagnosis through Active Test Selection and give each attribute a fixed privacy weight, set to its marginal uniqueness in the population. This is exactly what a sensible reviewer would build. On Pima diabetes ACTMED reports 68% accuracy against about 45% with all features, and near-perfect CKD accuracy with one or two tests per patient. We reimplement it with the same trained predictive model we use, rather than GPT-4o, so the only differences are the objective and the stopping rule. Its weights and KL threshold get the same tuning budget as our $\tau$ and $k$.
The third baseline is a static privacy-penalized feature subset. One globally chosen subset is selected by greedy forward selection with the same non-additive $R$ as a penalty. This tests whether per-patient adaptivity matters once privacy is priced correctly.
The fourth is input perturbation: each feature is released with noise calibrated to a local differential privacy level, which is the standard "add noise to everything" answer.
Two ablations isolate our ingredients. The first replaces action entropy with diagnosis entropy. The second replaces joint $R$ with the sum of marginal risks. The proposal beats the field if it matches full disclosure's action accuracy to within two points while releasing a tuple with lower re-identification risk than cost-augmented ACTMED at the same accuracy. If it only matches ACTMED, the non-additive risk term and the action-level stopping rule have not earned their place.
Why not just…?
Why not just add a privacy term to ACTMED's cost function? This is the real threat, and it is why that variant is our primary baseline. ACTMED's utility is additive in per-test cost (Timely Clinical Diagnosis through Active Test Selection). Re-identification risk is not additive: age, zip-derived region and a rare lab value may each be harmless and jointly unique. ACTMED's KL stopping rule also keeps going while diagnoses that share a treatment remain confused. The two ablations test each point separately. If the additive version comes within noise of ours on every dataset, we will report that the simpler system suffices.
Why not just de-identify the record and send it all? Scrubbing direct identifiers leaves the clinical attributes that enable linkage. An Anonymous Health Care System motivates its design precisely by the fact that longitudinal medical profiles re-identify patients. Rule-based scrubbers also remove clinically meaningful terms, as described in Annotation of a Large Clinical Entity Corpus. Our full-disclosure baseline is effectively "de-identified record," and its measured $R$ shows how much linkage risk remains.
Why not just pick one fixed minimal panel for everyone? A global subset must cover the hardest patients, so easy patients disclose more than they need. ACTMED's finding that three adaptively chosen features beat the full set on Pima and Hepatitis C suggests that per-patient choice matters. The static privacy-penalized baseline measures this directly under an identical risk metric.
Why not just add differential-privacy noise to every field? Noise that is small enough to protect a rare lab value is large enough to merge diagnoses that differ only in that value. That is exactly an artificial ambiguity zone, and the confusion map will show it. We keep this as a baseline rather than assume it loses.
Proposed Experiments and Decisive Checks
We use the benchmarks in the file that have structured, attribute-level data. These are the ACTMED datasets (CKD, Hepatitis C, Pima diabetes, and the OSCE-style multi-condition set from AgentClinic) and the six-class dermatology dataset with 366 samples and 34 features from Sensors 2023. The multi-condition and dermatology sets carry the main claim, because binary tasks cannot contain non-trivial ambiguity sets.
For each dataset we define action groupings with a clinician before running any model and freeze them. We then report three things. The first is action accuracy and diagnosis accuracy. The second is mean attributes disclosed and token round trips. The third is re-identification risk, measured as $k$-uniqueness of the released tuple within the dataset treated as the population, and alongside it an attribute-inference attack accuracy on held-out sensitive fields. Everything is reported as accuracy–risk curves swept over the budget.
The first decisive check is the confusion map. If most privacy-induced ambiguity on dermatology and the multi-condition set turns out to be action-relevant, the premise that action-level stopping saves disclosure fails, and we will say so.
The second decisive check is the head-to-head against cost-augmented ACTMED at matched action accuracy. Our method should release tuples with materially lower $k$-uniqueness. If it does not, the non-additive risk term is unnecessary.
The third check is calibration of returned ambiguity sets. When the budget is exhausted, the true diagnosis should fall inside the returned set at the nominal rate. This mirrors the engineering practice of handing an equivalence set to a human rather than guessing.
For the unstructured-text setting, DiReCT shows that GPT-4 reaches only 7.4% diagnostic accuracy without a knowledge graph (DiReCT). That result tells us text-level disclosure would first require reliable extraction of findings from notes. We therefore leave clinical notes to a follow-up, once the tabular results are in.
References
- Timely Clinical Diagnosis through Active Test Selection
- E-Tenon: An efficient privacy-preserving secure open data sharing scheme for EHR system
- Data-Driven Fault Isolation in Linear Time-Invariant Systems: A Subspace Classification Approach
- Fault Separation Based on An Excitation Operator With Application to a Quadrotor UAV
- Lightweight Multi-Class Support Vector Machine-Based Medical Diagnosis System with Privacy Preservation
- An Anonymous Health Care System
- Annotation of a Large Clinical Entity Corpus
- DiReCT: Diagnostic Reasoning for Clinical Notes via Large Language Models