How Often Should a Private Genomic Knowledge Base Update? Letting Clinician Behavior Decide

Clinics that use genetic information, for example to pick a drug dose from a patient's genes (pharmacogenomics), rely on shared reference tables. These tables record how common each drug-metabolism type is in each ancestry group, or which alternative drug has caused the fewest adverse events among slow metabolizers. When such tables are built from real patient records and published repeatedly, each release leaks a little more about the people in them. Differential privacy (DP), which adds calibrated random noise so that no single person's presence can be detected, protects them. The cost is that the more often you publish, the more noise each release needs. Publishing rarely keeps noise low, but the table goes stale while guidelines and cohorts change underneath it. So the operators of every such database face a real design question: how often to update.

Today that question would be answered by minimizing a statistical error. We argue it should instead be answered by what clinicians actually do with the table. The two kinds of error are not equally visible. Noise shows up as wider error bars or recommendations that wobble between releases. Staleness is silent: an outdated table looks exactly as confident as a fresh one. If clinicians lose trust over visible noise but act calmly on invisible staleness, a cadence tuned for statistical accuracy will look untrustworthy while silently doing more harm. Our proposal is to measure this trade-off in clinicians' decisions, not only in their stated attitudes. We then build the measured response into an extended technology acceptance model and use it to choose the update cadence.

Core Claim

Release cadence, privacy noise, and data staleness should be modeled as separate inputs to clinician acceptance. Staleness and noise of equal statistical harm will produce different drops in trust and different rates of correct action. As a result, the cadence that best supports correct decisions will differ from the cadence that minimizes error. That gap should be largest for "which option is best" releases, where the cost of continual DP is known to be steep.

Prior work does not address this. The Price of Differential Privacy under Continual Observation proves that answering "which item has the largest total" over $T$ releases costs error growing like $\tilde{\Omega}(T^{1/3})$ under approximate DP. Simple running counts cost only $O(\text{polylog}\,T)$. That paper never asks how humans respond to the resulting noise. Design and Implementation of PharMe is the closest system and touches all three directions. It runs guideline matching on the patient's phone, updates recommendations when co-medications change, and reports a System Usability Scale score of 82.86 with 28 users. However, it curates its guideline snippets by hand, gives no formal privacy guarantee for anything it aggregates, and never varies update cadence. Privacy-preserving genomic testing in the clinic supplies the behavioral motivation: 53% of physicians found encrypted genomic reports useful, but only 10% would change treatment. That is why our primary outcome is the decision a clinician makes, not their stated intention.

Research Program

The work has three parts.

Part 1: a semi-synthetic release stream. We simulate a pharmacogenomic registry in which patients enroll over a two-year horizon. Their metabolizer types per gene are sampled by ancestry from 1000 Genomes allele frequencies, the same public source used in Semantic Beacons. We inject timed drift events of two kinds. The first is guideline revisions modeled on the CPIC and FDA sources that PharMe curates. The second is enrollment shifts that change subgroup composition, so that a stale table is genuinely wrong. Two release types run on this stream.

Two release types. The first is a counting release: metabolizer-type frequencies by ancestry. We protect it with a tree-based continual counting mechanism, which adds noise to partial sums arranged in a binary tree so that error grows only logarithmically with the number of releases. The second is a selection release: the alternative drug with the lowest adverse-event rate among poor metabolizers in each subgroup. This is the task where continual DP is expensive. For a release frequency $f$ over a horizon $H$, the expected error of a published value can be written as

$$E(f) \approx \frac{\lambda}{2f} + \frac{c(fH)}{\varepsilon},$$

where $\lambda$ is the drift rate of the true value, $1/(2f)$ is the average age of a release, $\varepsilon$ is the total privacy budget, and $c(T)$ is the mechanism's noise growth over $T$ releases. For counting, $c(T)$ is roughly polylog $T$. For selection, it is roughly $T^{1/3}$ (both forms from the continual-observation lower bounds). This formula gives the statistician's optimal cadence, which is the reference point the behavioral study will either confirm or overturn.

Part 2: a factorial vignette experiment with practicing clinicians. Participants are physicians and pharmacists, roughly 300 to match the scale of the nurse survey in Artificial intelligence readiness … among newly qualified nurses. Each sees prescribing cases accompanied by a table drawn from the simulated stream. Within each participant we independently vary three things:

  • the amount of staleness error, crossed with whether data age is disclosed;
  • the amount of noise error, shown either as interval width or as recommendations that flip between consecutive releases;
  • the release type (counting or selection).

Staleness and noise levels are matched on statistical harm, meaning the same probability that the table leads to the wrong drug. That matching is what makes the asymmetry claim testable. For each case we record the prescribing choice and whether it matches the ground-truth correct action. After each block we administer short scales: perceived usefulness, ease of use, trust, and two new constructs, perceived freshness and perceived precision.

Part 3: modeling. We fit a structural equation model linking the manipulated error sources, through the new and standard acceptance constructs, to the observed choices. We then plug the fitted response curves back into the stream simulator. This yields a "behavioral optimal cadence": the update frequency that maximizes the expected rate of correct clinical action at a fixed privacy budget.

Baselines

Naive reference. A single DP-protected snapshot is published once and never updated. This is effectively what the HIV clinic system did with its static risk mappings (Privacy-preserving genomic testing in the clinic). It bounds how bad staleness gets.

Reviewer default. Quarterly batch releases, each given an equal share of the budget. Its cadence comes from convention, not optimization. It is run on the identical stream with the identical total $\varepsilon$.

Strongest existing method: the error-minimizing cadence. For counting, this is the tree-based continual release analyzed in The Price of Differential Privacy under Continual Observation, with $f$ chosen to minimize $E(f)$. For selection, it is the corresponding continual selection mechanism. This is the strongest competitor because it is exactly what a careful privacy engineer would deploy. For the counting stream we also evaluate a local-DP variant using LOLOHA from Frequency Estimation of Evolving Data Under Local Differential Privacy. That paper reports reducing privacy budget consumption by 2 to 25 orders of magnitude relative to RAPPOR on UCI and Census streams. We found no reported result on Lacuna for any continual-release mechanism on a pharmacogenomic or clinical genomic stream. All mechanisms share the same stream, the same total budget, and the same grid of candidate cadences, so none gets extra tuning.

What counts as beating it. The behavioral optimal cadence must yield a higher rate of correct prescribing choices than the error-minimizing cadence, at the same $\varepsilon$, on held-out simulated cases. The improvement must be supported by confidence intervals from a participant-level bootstrap.

Measurement-model baseline. On the measurement side, the baseline is a standard UTAUT/TAM model with the trust and privacy-concern extensions used in the Dschang cervical-screening study (Acceptability of artificial intelligence for cervical cancer screening in Dschang, Cameroon). It is fit to the same participants.

Ablations. The key ablation merges perceived freshness and perceived precision into one "data quality" construct. If the merged model predicts held-out choices as well as the split model, the core claim fails. A second ablation removes data-age disclosure. It tests whether simply showing a table's age closes the staleness blind spot.

Why not just…?

Why not just minimize statistical error and ignore perception? If clinician choices track expected error, then the behavioral optimal cadence will coincide with the statistical one, and the simple approach is sufficient. This is a real possibility. The experiment is designed so that this outcome is reportable and decisive, not hidden. The reason to doubt it in advance is that stated utility and action already diverge sharply in genomic reporting (53% versus 10% in the HIV clinic study). There is also evidence that perceived usefulness can fail to predict readiness at all, as in the nurse study.

Why not just release in fixed quarterly batches, since guidelines change slowly? Quarterly batching is one of our baselines, so the data will show whether it is good enough. The argument against assuming it is that the content that drifts includes more than slow-moving guidelines. Cohort composition also changes, and changes in which subgroups are well represented are fast and invisible. PharMe's handling of co-medication changes, which alter a patient's effective metabolizer type, likewise shows that the clinical inputs are not static.

Why not just add no noise, given that realistic re-identification risk is low? A decade of research on genetic privacy finds that 95th-percentile re-identification risk in All of Us stays below regulatory thresholds, and that the public weighs clinical utility above abstract privacy risk. This is a serious challenge. Our response is to include a no-noise arm, so the study directly measures the behavioral cost of DP at each cadence. That cost is the number a policymaker needs in order to decide whether DP is worth paying for. The same review reports heightened distrust among marginalized groups, and small ancestry-by-rare-genotype cells are exactly where both privacy risk and noise concentrate. We therefore report effects separately for small subgroups.

Why not just use a standard UTAUT survey? The measurement-model baseline does exactly this. The split-construct ablation tests whether the new constructs improve prediction of held-out choices, not just model fit. If they do not, we drop them.

Proposed Experiments and Decisive Checks

All releases come from the semi-synthetic stream, at privacy budgets of $\varepsilon \in \{0.5, 1, 4\}$ plus a no-noise arm. Cadences range from monthly to annual. Stream code reuses residual-based marginal reconstruction from bcmullins/efficient-marginal-reconstruction for multi-way subgroup tables. Primary metrics are the correct-action rate and the action rate on vignettes. Secondary metrics are trust and perceived usefulness scores, and prediction of held-out choices measured by log-loss and AUC.

Four results would settle the proposal:

  • Asymmetry. At matched statistical harm, does noise lower trust significantly more than staleness, while staleness lowers correct-action rate at least as much? This is the asymmetry hypothesis.
  • Cadence gap. Does the behavioral optimal cadence differ from the error-minimizing cadence, and is the gap larger for selection releases than for counting releases, as the $T^{1/3}$ versus polylog growth predicts?
  • Disclosure. Does showing data age remove the staleness blind spot? If so, disclosure is a cheap fix, and the cadence question shrinks back to a statistical one.
  • Constructs. Do the split freshness and precision constructs predict actual choices better than standard UTAUT? This would show they add something beyond stated intention.

A null result on the first two checks would show that error-minimizing cadence is safe to deploy. That outcome is itself useful, and we would report it as such.