Federated, Event-Conditional Conformal Bands for Intraoperative Blood Pressure Forecasts

Deep models can now forecast a surgical patient's mean arterial pressure (MAP) several minutes ahead from ordinary 30-second bedside vital signs. None of them says when it might be wrong. Adding uncertainty in the usual way would not fix the problem that matters. A prediction interval that covers the true pressure 90% of the time on average earns almost all of that coverage during the long, stable stretches of surgery. It can miss the rare minutes when pressure actually falls below the hypotension threshold, and those are the only minutes that matter clinically.

This note proposes a different target. We want intervals whose coverage is guaranteed conditional on the patient actually being hypotensive, so that "the band dips below 65 mmHg" becomes an alarm with a provable lower bound on how many hypotensive episodes it catches. Calibrating such a guarantee needs many hypotensive examples. Any single hospital has few of them and cannot share raw waveforms. The proposal therefore federates only the calibration step: sites exchange compressed summaries of their forecast errors on hypotensive minutes, never patient data.

Core Claim

We can obtain a finite-sample guarantee of the form "with probability at least $1-\alpha$, every hypotensive minute of a new surgery falls inside the forecast band." This is a surgery-level recall guarantee for the alarm. We get it by combining three pieces. The first is outcome-conditional conformal prediction, a distribution-free method that turns any forecaster's past errors into intervals with guaranteed coverage. The second is calibration at the level of whole surgeries rather than individual time steps. The third is federated pooling of quantile sketches of these errors across hospitals.

The federation is what makes the guarantee usable. A small hospital with a few dozen hypotensive surgeries would otherwise get bands so wide that the alarm fires constantly. The claim is falsifiable. It predicts that, at matched guaranteed recall, the federated bands produce fewer false alarms than site-local calibration at small sites, and that marginal federated conformal under-covers hypotensive minutes.

Where This Sits

The forecasting side supplies the base model and nothing more:

  • CMA forecasts continuous MAP trajectories from 30-second vital signs so clinicians can choose thresholds at runtime. It reports 94.1% accuracy at a 5-minute lead and a 4.48 mmHg MAE at 15 minutes on Tongji data.
  • HMF adds trend/seasonal decomposition and instance normalization for drifting baselines. Its companion paper (Hybrid Multi-Factor Network) documents a sharp drop on elderly patients.

Both are centralized and output only point estimates. We reuse them unchanged as forecasters.

CSA-TTA motivates the rare-event framing. It shows that models adapted on a patient's stable segments smooth over crashes. Its remedy, however, is a central bank of raw historical segments and a heuristic alarm trigger, which is exactly what a multi-hospital privacy setting forbids.

The uncertainty side supplies the method and its limits:

  • Federated Conformal Predictors (FCP) supplies the machinery. Clients send T-Digest sketches of their nonconformity scores, which are per-example error magnitudes. The server's quantile gives marginal coverage between $1-\alpha$ and $1-\alpha+K/(N+K)$, assuming exchangeability within each client. FCP is marginal and was built for classification. We make it outcome-conditional, apply it to regression, and calibrate at the surgery level so that dependence within a time series does not break exchangeability.
  • Bayesian federated methods, namely FedIVON, FedWBA and FedGVI, constrain the design. FedGVI's own motivation is that Bayesian uncertainty miscalibrates under misspecification, and artifact-laden vitals are misspecified almost by definition. We therefore treat Bayesian credible bands as a competitor rather than the mechanism.
  • The EHR federated learning review motivates the setting. It finds no federated work on high-resolution intraoperative series and little principled uncertainty in engineering-style federated learning.

Mechanism

Let $\hat{y}_{t+h}$ be the forecaster's MAP prediction $h$ minutes ahead, and let $s = |y_{t+h}-\hat{y}_{t+h}|/\hat\sigma_{t}$ be the nonconformity score, where $\hat\sigma_t$ is a scale estimate. That scale estimate can come from a quantile head or from recent residual magnitude. Define the stratum $g(y)=1$ if $y<\tau$, with $\tau$ the hypotension threshold (65 mmHg by default), and $g(y)=0$ otherwise.

For each calibration surgery containing hypotension, a site records a single score: the maximum of $s$ over that surgery's hypotensive minutes. It then sends a T-Digest sketch of these per-surgery scores to the server. The server merges the sketches into a quantile $q_1$. The same procedure gives $q_0$ for the non-hypotensive stratum.

The prediction set is

$$C(x_t) = \{\, y : |y-\hat{y}_{t+h}| \le q_{g(y)}\,\hat\sigma_t \,\},$$

which is in general a union of two intervals, one on each side of $\tau$. Because the stratum is defined by the candidate $y$, the set remains valid even though the future outcome is unknown. The alarm fires when $C(x_t)$ contains any value below $\tau$.

Using the per-surgery maximum yields the event-level statement. Under partial exchangeability across hypotensive surgeries, the FCP argument gives

$$\Pr\big(\text{all hypotensive minutes of a new surgery lie in } C \big) \ge 1-\alpha .$$

This amounts to a surgery-level recall floor of $1-\alpha$ for the alarm. The upper slack is $K/(N_1+K)$, where $K$ is the number of sites and $N_1$ the number of pooled hypotensive surgeries. That slack is the quantitative reason to federate: with $N_1$ in the tens at one site the bound is loose and $q_1$ is noisy, and with hundreds pooled it tightens.

Two refinements address the file's known failure modes:

  • Age crossing. We cross the outcome stratum with an age band, which is known at test time, to target the elderly gap.
  • Site-specific shrinkage. A site uses a weighted blend of its local and the global sketch quantiles. The weight grows with its local $N_1$. This keeps coverage from relying on FCP's fixed-mixture assumption when a site's error distribution differs.

For drift within a surgery we test an online adaptive update of $q$. We report it as an add-on, not as the source of the guarantee.

Research Program

Base forecaster. We train HMF and CMA with ordinary federated averaging on 30-second vitals, so that deployment matches standard bedside monitors. Training the forecaster is not the contribution. We fix the forecaster across all uncertainty methods so that differences come only from calibration.

Federation. We build a simulated federation from VitalDB (6,388 cases) by partitioning surgeries into sites by surgical category and age composition from the recorded clinical information. We deliberately make some sites small, with 20 to 50 hypotensive surgeries. CH-OPBP (1,083 patients, downsampled) serves as a distinct hospital. One site is fully held out to test transfer to a new hospital.

Theory. We will state the coverage argument under surgery-level partial exchangeability. We will quantify how the bound degrades when the test site's mixture weight is misestimated, which is the known weak point of FCP.

Baselines

Naive reference. The fixed forecaster alarms when its point forecast crosses 65 mmHg, with no uncertainty. The relevant published figures are HMF's recall of 67.98% and AUC of 0.7352 on CH-OPBP, CMA's 94.1% accuracy at 5 minutes on Tongji data, and CSA-TTA's zero-shot recall gain of +7.46% on VitalDB. No reported result found on Lacuna for any calibrated uncertainty method on intraoperative MAP forecasting.

Tuned point-forecast threshold. We pick, per site or on pooled sketches, the alarm threshold on the point forecast that achieves a target recall. This is the strongest simple competitor and is discussed below.

Marginal federated conformal. FedAvg forecaster plus FCP intervals exactly as in FCP, with no outcome stratification. For reference, FCP reports prediction sets 1.2× the centralized size on CIFAR-100 under heterogeneity, against 3.8× for plain FedAvg-based calibration. That is a classification number, not directly comparable here.

Site-local outcome-conditional conformal. The same method without federation. This ablation isolates the value of pooling.

Bayesian credible bands. FedIVON credible bands from a variational version of the same architecture. The Bayesian competitor gets a hyperparameter budget matched to our calibration-split tuning.

Centralized pooled calibration. An upper-bound oracle that pools all calibration surgeries in one place.

Fairness and success criteria. Every method uses the same train, calibration and test surgeries. Every method is compared at matched nominal level ($\alpha = 0.1$ and $0.05$). We beat a baseline if we reach the target surgery-level recall while it does not, or if both reach it and we produce fewer false-alarm minutes per surgery and a smaller mean band width. The gains must hold on small and held-out sites specifically. Further ablations remove the per-surgery maximum (time-step calibration instead), remove the age crossing, and remove shrinkage.

Why not just…?

Why not just tune the alarm threshold on the point forecast until recall hits the target? This is the real threat. Our alarm is close to a thresholded score, and at a large site a tuned threshold may match it. The differences are three:

  • Conformal gives a finite-sample guarantee rather than an empirical estimate that can overfit a few dozen events.
  • It yields a band valid at clinician-chosen thresholds on a precomputed grid (60, 65 and 70 mmHg), in the spirit of CMA's runtime flexibility.
  • The federated version pools rare events that a small site cannot.

If federated threshold tuning over shared sketches matches us on held-out recall variance and false alarms, we will report that the value lies in the pooling and the guarantee, not in the band itself.

Why not just calibrate conformal intervals locally at each hospital? Nothing prevents it. The objection is that with few hypotensive surgeries the stratum quantile $q_1$ is either undefined (infinite bands) or highly variable, and the $K/(N+K)$-type slack is large. The site-size sweep directly measures where local calibration becomes competitive. If it does so at realistic sizes, the federated component is unnecessary and we will say so.

Why not just use marginal federated conformal prediction? Marginal coverage is dominated by stable minutes. We expect it to reach 90% overall while under-covering hypotensive minutes, the same stability bias that CSA-TTA reports for adaptation. The stratified coverage breakdown will show whether this happens.

Why not just use Bayesian federated learning such as FedIVON? It gives uncertainty with no coverage guarantee. FedGVI itself argues that such uncertainty degrades under misspecification and noise. We test FedIVON directly: if its credible bands attain event-conditional coverage, the conformal layer adds little.

Why not just pool the data centrally? Legal barriers to pooling are documented in the review. The centralized oracle tells us how much the federated version gives up relative to pooling.

Decisive Checks

Coverage on hypotensive minutes and surgeries. The primary result is empirical surgery-level coverage of hypotensive minutes on held-out sites. We report it separately for elderly and non-elderly patients and plot it against nominal $1-\alpha$. Method-level expectations are:

  • Marginal FCP should fall visibly below target.
  • Our method should stay at or above target.
  • Site-local conformal should meet the target only with much wider bands at small sites.

Operating quality. Secondary results are false-alarm minutes per surgery at matched guaranteed recall, and median band width at 5, 10 and 15-minute horizons.

Robustness to site shift. To stress the mixture assumption, we hold out an entire surgical category and also inject artifact segments into one site's data (flat-lines and flush spikes). We then measure whether shrinkage preserves coverage.

Latency. We report per-step inference latency against CSA-TTA's 1.7 to 6.6 seconds per update. Our method adds only a lookup of the two precomputed quantiles to the forecaster's forward pass.

What would falsify the proposal. The proposal fails if the federated bands deliver no false-alarm reduction over site-local calibration at 20 to 50 hypotensive surgeries per site. It also fails if the per-surgery-maximum construction produces bands too wide for a usable alarm.