Health privacy law decides what to protect one data stream at a time. A hospital record is protected, while a calendar entry, a step count, a pharmacy visit in a location trace, or a music-listening history usually is not. The most worrying inferences, however, come from combinations. A phone's mobility log, its screen-unlock pattern and its sleep estimate can each be uninformative about depression on their own and still reveal it together. We propose to measure this directly. The plan is to build an audit that, for a given sensitive health condition, finds the minimal sets of data streams whose joint release lets an attacker infer the condition clearly better than any smaller subset can. The output is a "leakage hypergraph": each node is a data stream, and each hyperedge is a group of streams that is jointly dangerous even though no part of it is.
The second half of the idea is a correction the audit cannot skip. Any such audit has to score the attacker against a label, and in behavioral health that label is almost always a questionnaire answer such as a depression score or a stress rating. Self-reports are distorted by recall, by the wish to look well, and by guessing what the study wants. An attacker scored against a distorted label will appear weaker or stronger than it really is against the person's actual state. So the audit must report leakage together with bounds that account for how biased the label is, not a single raw number.
Core Claim
Inference risk from unregulated health data sits in specific small groups of streams. Audits that rank streams one by one, as current regulatory categories do, cannot see these groups, and pairwise audits miss part of them. These groups can be found efficiently by adapting the sufficiency-and-necessity subset search used in clinical hypergraph models. Because the scoring labels are self-reports, the size of the leakage should be reported as an interval derived from a measurement-error model, not as a raw accuracy.
The motivating evidence comes from the privacy literature. Physiological Data: Challenges for Privacy and Ethics defines "auxiliary use": passive physiological signals are fused with non-physiological footprints such as GPS or browsing history to infer unshared mental states. The paper is conceptual and offers no way to measure this. Shadow Health-Related Data: Definition, Categorization, and User Perspectives shows that health traces outside the clinic are nearly universal, with 95% of its 300 respondents using search engines for health queries and 87% putting appointments in general calendars. It treats each channel as a separate bucket and relies entirely on self-report. Wearable Activity Trackers: A Survey on Utility, Privacy, and Security documents that US statutes do not classify tracker data as protected health information. None of these three works computes which combinations matter, and computing that is what this proposal does.
Research Program
The central quantity is the synergy of a set of streams. Let $L(S)$ be the leakage achievable from stream set $S$, measured as the held-out AUROC of the best attacker trained only on $S$, where held-out means evaluated on people not used for training. The synergy of $S$ is defined as
$$\Delta(S) = L(S) - \max_{T \subsetneq S} L(T),$$
which is the gain from having all of $S$ over the best proper subset of it. A leakage hyperedge is a set $S$ for which $\Delta(S)$ exceeds a threshold $\tau$, chosen by a permutation test, while every subset of $S$ falls below it. Defining hyperedges this way makes them genuinely indecomposable groups, in the sense used by hypergraph learning work such as Counterfactual and Factual Reasoning over Hypergraphs for Interpretable Clinical Predictions on EHR. The difference is that here a hyperedge is an output of the audit, not an input to a predictor.
For coarse streams, for example roughly ten sensing modalities, all subsets up to size three can be enumerated. The real difficulty appears at the finer level where regulation and consent actually operate: individual feature groups such as "night-time WiFi location changes" or "conversation duration." At that granularity the number of subsets explodes, which the hypergraph literature names as an open problem in constructing hyperedges across heterogeneous streams.
Our search method borrows from CACHE, which uses Gumbel-max relaxations to extract a minimal subset of medical codes that is sufficient, meaning it reproduces the prediction, and necessary, meaning that removing it destroys the prediction. We apply the same mask objective at the level of data streams, using a joint attacker model that accepts arbitrary stream masks. We then add a repulsion term, adapted from the prior in Disentangling Latent Risk Pathways via Bayesian Hypergraph Inference, which pushes successive masks apart so that the search returns several distinct leakage groups rather than many near-copies of the same one. Every candidate set the search proposes is re-verified by retraining an attacker from scratch on exactly that set. The reported $\Delta(S)$ therefore never depends on the search model's own internal attributions.
The self-report correction rests on a simple observation. Scored against a label that flips at random, AUROC is pulled toward 0.5, so audits understate risk. Scored against a label whose bias tracks behavior, AUROC can move in either direction. Social desirability is one such case: Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders reports systematic underreporting, and heavy phone users may underreport both phone use and distress. To handle this, we model the reported label as a noisy version of the true state, with misclassification rates that may depend on a small set of observed covariates. Using repeated measurements of the same construct within a person, we estimate the plausible range of those rates and then report $L(S)$ and $\Delta(S)$ as intervals over that range.
Repeated measurements constrain reliability, not validity. The correction therefore produces bounds, not point identification, and we will say so. The useful finding would be that the ordering of hyperedges stays stable across the bounds even when their exact sizes do not.
Baselines
The naive reference point is the audit that current regulation implies: rank each stream by the leakage it allows alone and protect the top ones. It is run with the same attacker family, the same leave-people-out splits and the same tuning budget as every other method. If our hyperedges simply coincide with the top single streams, the proposal has nothing to add.
The second baseline is the pairwise audit, which enumerates all stream pairs. It is the audit-side analogue of building a pairwise graph by linking every member of a group, and it plays the role that graph attention baselines played in CACHE. There, hypergraph modeling reported a 3.2% AUROC gain and a 7.5% AUPR gain over the best baselines on MIMIC-III. Separately, the BHPI paper reported a +1.2 AUC gain over independent logistic regression on rare diseases. To beat the pairwise audit, we must find triples or larger sets with significant $\Delta(S)$ that no pair captures.
The strongest existing attacker on our main dataset is the spatiotemporal graph neural network of Monitoring Behavioral Changes Using Spatiotemporal Graphs: A Case Study on the StudentLife Dataset. It reports stress-classification F1 of 65.4 ± 3.1, accuracy of 67.1 ± 2.7, and depression-score RMSE of 3.5 ± 3.1 using all features. We will reproduce it as the full-access attacker. The audit is only credible if the attackers it retrains on subsets reach that level when given all streams. A weaker attacker would make every synergy estimate an underestimate.
The strongest method competitor for finding hyperedges is to train one joint model and compute pairwise and higher-order interaction attributions from it, using Shapley-style interaction indices. It runs on the same trained model and data, and its top-ranked sets are verified by the same retraining step. We found no reported result on Lacuna for interaction attribution as a leakage audit. Our search beats it if it recovers more planted hyperedges at equal retraining cost, and if the sets it finds have larger verified $\Delta(S)$.
Two ablations isolate the key ingredients. The first removes the repulsion term to test whether diversity matters. The second removes the label correction, reporting raw AUROC against the self-report, to test whether the correction changes which sets are flagged and not merely their sizes.
Why not just…?
Why not just train one model on all streams and compare it to single-stream models? This check is cheap, and it is our first experiment, because it decides whether the project is worth continuing. If the gap between the full-access attacker and the best single stream is within noise on StudentLife, combination leakage is not the main risk there, and the proposal should stop. If the gap is large, the check still says nothing about which minimal groups produce it. Consent forms, data-broker rules and withholding policies need exactly that information. Experiment four tests whether knowing the groups buys more protection per withheld stream.
Why not just read interactions off a trained model with attribution methods? This alternative is a real threat. Attribution describes what one fitted model relied on. When streams are redundant, that model may put weight on one stream of a redundant pair, and the audit would then miss a different group an attacker could exploit. Interaction indices beyond pairs are also expensive and high-variance at a sample size of 48 people. We do not argue the point. The planted-hyperedge experiment compares the two methods directly, and if attribution matches our search, we will recommend it as the cheaper tool.
Why not just fix the labels with psychometrics? The cross-lagged study in Longitudinal relationships among problematic mobile phone use, bedtime procrastination, sleep quality and depressive symptoms in Chinese college students: a cross-lagged panel analysis established measurement invariance with CFI = 0.999 and RMSEA = 0.032, yet still acknowledged information bias it could not remove. Invariance shows that a scale means the same thing across time or groups, not that it is accurate. We use psychometric structure as one input to the bias bounds, not as a replacement for them.
Why not just add local differential privacy noise to every stream? Adding random noise to each stream before release is a defense, not an audit. Applying it uniformly spends the privacy budget equally on harmless and dangerous streams. Graph-oriented work such as Privacy-Preserving Graph Embedding based on Local Differential Privacy protects graph structure, not health inference from groups of streams. We treat uniform noise as a comparison in experiment four: does concentrating noise on hyperedge members achieve the same leakage reduction at lower utility cost?
Proposed Experiments
The main testbed is StudentLife: 48 students over 10 weeks, with GPS, WiFi, accelerometer, conversation and phone-usage streams, daily stress EMAs and pre/post PHQ-9 scores, as used in the STGNN paper. All evaluation holds out whole people, and every $\Delta(S)$ is accompanied by a permutation-test p-value, because the small sample makes false hyperedges the main risk. A second, larger setting is UK Biobank as used in the BHPI paper. There, the 71 baseline factors are grouped into lifestyle-questionnaire, demographic and biomarker streams and serve as stand-ins for consumer footprints, with the 66 disease outcomes as targets. Where the cohort provides both self-reported and record-based versions of a condition, we audit against both. The Last.fm tags of Tag2Risk: Harnessing Social Music Tags for Characterizing Depression Risk, against which an SVM reached 66.4% accuracy on K10 labels, provide a single-platform check where the tag groups act as fine-grained streams.
Experiment one tests whether synergy exists. On StudentLife at modality level, we compute $L$ for every stream set up to size three for stress and depression and report the distribution of $\Delta(S)$. Experiment two tests whether the search scales. We plant synthetic sensitive labels that depend on known groups of real features, varying group size, signal strength and redundancy, so that the true hyperedges are known. We then measure recall and precision of planted hyperedges for exhaustive search where it is feasible, greedy forward selection, interaction attribution, and our search with and without repulsion, all at matched retraining budgets.
Experiment three tests the label correction. In semi-synthetic runs, we corrupt clean planted labels with known random and behavior-dependent misclassification. We check that raw audits mis-rank the hyperedges and that the corrected intervals cover the clean-label leakage. On real StudentLife labels, we report how wide the intervals are when bounded by within-person agreement between daily stress EMAs and PHQ-9.
Experiment four is the decisive test for policy relevance. For a fixed number of withheld streams, we compare the drop in full-access attacker AUROC under three policies: hyperedge-guided withholding, single-stream-ranked withholding, and uniform noise. The proposal is supported if hyperedge-guided withholding yields a clearly larger drop at equal cost and the advantage survives the label-bias bounds. It fails if single-stream ranking does as well.
References
- Physiological Data: Challenges for Privacy and Ethics
- Shadow Health-Related Data: Definition, Categorization, and User Perspectives
- Wearable Activity Trackers: A Survey on Utility, Privacy, and Security
- Counterfactual and Factual Reasoning over Hypergraphs for Interpretable Clinical Predictions on EHR
- Disentangling Latent Risk Pathways via Bayesian Hypergraph Inference
- Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders
- Monitoring Behavioral Changes Using Spatiotemporal Graphs: A Case Study on the StudentLife Dataset
- Longitudinal relationships among problematic mobile phone use, bedtime procrastination, sleep quality and depressive symptoms in Chinese college students: a cross-lagged panel analysis
- Privacy-Preserving Graph Embedding based on Local Differential Privacy
- Tag2Risk: Harnessing Social Music Tags for Characterizing Depression Risk