Models that forecast a patient's next few hours of vital signs, such as heart rate, blood pressure and oxygen saturation, are almost always trained with squared error. Squared error treats every mistake of the same size as equally bad. At the bedside they are not equally bad. Suppose a patient's mean arterial pressure (MAP) is about to fall to 58 mmHg, below the 65 mmHg level where clinicians act. A forecast of 66 is "one unit wrong" about the threshold but tells staff that nothing needs doing. A forecast of 50 is further off numerically but prompts the same response. Robotics and autonomous driving ran into the same problem with trajectory forecasts that clip obstacles. Their answer was collision-aware losses, meaning training penalties that grow only when a predicted path enters a forbidden region.
This note proposes the clinical counterpart. We treat danger thresholds on vital signs as obstacles and train forecasters with a penalty for false reassurance: predicting the safe side of a threshold when the patient actually crosses it. The central claim is testable. We argue that false reassurance is where cross-hospital and cross-population failure concentrates, for a specific mechanical reason. Clinical recordings are rounded toward "normal" values, and how much they are rounded depends on the hospital, the ward and the patient's demographics. A model trained with squared error on this data therefore learns an optimism that differs by site and by group. A penalty targeted at boundary crossings should shrink both the external-cohort gap and the subgroup gap in false-reassurance rates. It should do this better than either generic asymmetric losses or generic fairness reweighting.
Core Claim
Three findings in the file motivate the idea. The quality of vital signs measurements in electronic medical records varies by hospital, specialty, and patient demographics analyzed about 4.3 million records from Oxford University Hospitals. It found that 15.1% of temperatures were recorded as exactly 36.0°C, against 4.9% expected. Rounding was stronger in older and male patients, stronger after a previous normal reading, and varied by specialty. The paper is purely observational and proposes no modeling remedy. Neuro-Symbolic World Models for Counter-factual Safety in Automated Driving showed that world models trained on mostly-safe demonstrations become optimistic "scene completers." Reconstruction loss does not register their false-safe failures, and adding rule-based energy penalties cut false-safe predictions from 27.9% to under 1% on simulated driving. MIMIC-IF found that model AUROC is inversely related to a demographic group's baseline mortality. In other words, models perform worst on the patients most likely to be in danger.
Taken together, these suggest a causal chain. Normalcy-biased recording plus squared-error training produces forecasts pulled toward the safe interior. The size of that pull depends on local recording habits and on patient group. So a model's false-reassurance rate should rise when it moves to a new hospital and should be highest in the groups whose recordings are most rounded. We claim that a boundary-crossing penalty, calibrated from physiology and not tuned to any one site, reduces this failure mode where it is worst. We also claim it is the boundary structure that matters, not asymmetry as such.
The loss adds a crossing term to a standard forecasting objective. Let $y_t$ be the true value at horizon $t$ and $\hat y_t$ the forecast. Let $D$ be a clinically defined danger region, for example MAP below 65 or SpO2 below 90. Let $d(\hat y_t, D)$ be the distance from the forecast to $D$, which is zero if the forecast is already inside $D$. The objective is
$$ \mathcal{L} = \sum_t (\hat y_t - y_t)^2 + \lambda \sum_t \mathbb{1}[y_t \in D]\, d(\hat y_t, D)^2 , $$
where $\lambda$ is a single weight. For a probabilistic forecaster, the second term is replaced by the predicted probability mass that lies outside $D$ when the truth lies inside it. This mirrors how Social-DPF penalizes overlap between predicted densities instead of distances between points. Because $D$ can be two-sided (hypotension below, hypertension above), the direction of the penalty flips depending on where the patient is. A single uniformly asymmetric loss cannot represent that.
Research Program
The method comes from driving. Three papers supply it: the conflict loss in VADv2, which penalizes planned trajectories that intersect obstacles or road edges; the off-road loss in Trajectory Prediction for Autonomous Driving based on Multi-Head Attention with Joint Agent-Map Representation, which penalizes forecasts that leave the drivable area; and the collision-aware remediation studied in SafeShift. SafeShift is the closest prior intersection. It showed that the Motion Transformer's predicted collision rate on the Waymo Open Motion Dataset rose from 0.044 in-distribution to 0.100 on a safety-critical held-out split, and that collision-aware losses cut predicted collisions by about 10% on average, at a small cost in displacement error. We borrow its evaluation logic: hold out the dangerous cases and measure safety-specific error, not average error. We differ from it in three ways. The shift we study is across institutions and populations, not across scenario types. The "obstacle" is a physiological threshold, not another agent. And we test a mechanism, recording bias, that has no counterpart in driving logs.
On the clinical side, Neural Control System for Continuous Glucose Monitoring and Maintenance already puts physiological-limit penalties into a differentiable loss. It does so for a controller on a synthetic linear model, and its author notes it has never been tested on real multi-site patients. We move in the opposite direction: a forecaster, real records, several sites, no simulator. Prior sepsis work on cross-site data is our main source of evidence that the gap is real. Early Prediction of Sepsis from Clinical Data: the PhysioNet/Computing in Cardiology Challenge 2019 reports marked degradation on the withheld hospital. Representation Learning for Early Sepsis Prediction reports that a reconstruction-trained LSTM autoencoder degraded substantially on the external site, which is direct evidence that unconstrained squared-error representations do not transfer. Neither work examined which direction the errors fell in.
The program has three parts.
The first is a measurement study. On each available cohort we estimate terminal-digit preference and rounding toward normal values, by site and by subgroup, using the approach of Jackson et al. We then test whether a squared-error forecaster's signed bias near thresholds tracks that rounding.
The second is the loss itself. It is applied to fixed forecasting backbones, a GRU and a small Transformer over hourly binned vitals, at horizons of 1 to 6 hours. Thresholds are taken from standard clinical criteria and fixed before any external data is inspected.
The third is locked-model external validation. Every model is frozen before it touches an external cohort. This follows the transportability framing of All models are local, which documents the Epic Sepsis Model falling from a marketed AUROC of 0.73–0.83 to 0.60–0.64 externally.
The data plan uses MIMIC-IV as the derivation cohort, processed with the open pipeline from An Extensive Data Processing Pipeline for MIMIC-IV. MIMIC-IV carries race and insurance, so it supports subgroup audits. The public PhysioNet 2019 hospital systems A and B serve as two external cohorts, restricted to the vital signs they share with MIMIC-IV. Train-on-A, test-on-B and the reverse give a second, purely cross-institutional test. PhysioNet 2019 exposes only age and sex, so race-level audits are internal to MIMIC-IV and external audits use age and sex. We will state this asymmetry openly rather than paper over it.
The primary metric is the false-reassurance rate. It is the fraction of true threshold crossings at horizon $t$ for which the forecast (for probabilistic models, the forecast at a fixed probability level) stays on the safe side, a direct analogue of the NeSyWM false-safe rate. Its complement, the false-alarm rate, is reported alongside it, together with MAE and interval coverage. For subgroup and site gaps we report the maximum pairwise difference in false-reassurance rate, with patient-level bootstrap intervals. As a downstream check we feed forecasts into sepsis prediction and report AUROC and the Challenge utility score, so that any gain in safety can be weighed against clinical utility.
Baselines
The naive reference is a squared-error forecaster with the same backbone, data and training budget. This is the "do nothing" point, and it is the model the rounding hypothesis predicts will be optimistic. There is no reported result on Lacuna for false-reassurance rates of vital-sign forecasters on MIMIC-IV or PhysioNet 2019, so every forecasting number here will be produced in this study under a shared protocol.
The strongest downstream competitor is the SailOcean pipeline from Early Prediction of Sepsis Using Multi-Feature Fusion Based XGBoost Learning and Bayesian Optimization. It reports a utility score of 0.364 on the full hidden 24,819-patient set, at a threshold of 0.522, and a local AUROC of 0.847. It is the strongest reported cross-site sepsis system in the file. We cannot reproduce the hidden set, so we will rerun its published feature recipe on our A/B splits. On the forecasting-to-sepsis pathway, "beating it" means matching its utility on the held-out hospital while lowering false reassurance. We do not expect to exceed it on utility.
Three further baselines are run on the same backbone. Quantile (pinball) regression with a two-sided prediction interval, where the interval edge nearest the relevant threshold is used, tests whether asymmetry alone explains any gain. A multi-threshold cost-sensitive classifier head, which directly predicts "will cross $D$ within $t$ hours" with weighted cross-entropy, tests whether forecasting is needed at all. Group distributionally robust optimization over MIMIC-IV subgroups, a standard method that minimizes the worst group's loss, tests whether generic fairness training closes the subgroup gap. All models get the same hyperparameter search budget, 30 trials each, tuned only on a MIMIC-IV validation fold. No external data is used for tuning.
The key ablation replaces the crossing penalty with a symmetric reweighting of errors near the boundary, which has the same total weight but no directional term. This isolates whether direction matters or whether the model is simply paying more attention to the band near the threshold. A second ablation trains on vitals that have been "de-rounded" by jittering values that fall on round numbers, to test the rounding mechanism directly.
Why not just…?
Why not just use quantile regression and read off a conservative quantile? This is the most serious threat. A lower-quantile forecast of MAP is automatically pessimistic about hypotension. Its weakness is that it is pessimistic everywhere and in one fixed direction. It inflates false alarms for patients far from any threshold, and it cannot handle a variable with danger on both sides without a separate model for each side. If the two-sided interval baseline matches our false-reassurance rate at an equal false-alarm rate across all external cohorts, we will conclude that the crossing structure adds nothing and report the proposal as a negative result. The comparison is run at matched false-alarm rates for exactly this reason.
Why not just train a classifier for threshold crossing? Cost-sensitive classification optimizes the decision we care about. For a single fixed threshold it may well be enough. The argument for forecasting is that one continuous forecast serves many thresholds, including patient-specific targets, without retraining. That is an argument about convenience, not accuracy. We therefore evaluate on thresholds held out from training, for example MAP below 60 when training used below 65. If the classifier head matches the forecaster even there, the continuous forecast has no advantage beyond being a single model.
Why not just apply group DRO or reweighting for fairness? Group-level reweighting raises the loss on underperforming groups but is blind to which errors those groups suffer. The MIMIC-IF inverse correlation between baseline mortality and AUROC suggests the deficit is specifically on dangerous cases. Group DRO also cannot anticipate a new hospital's recording habits. The decisive comparison is the external subgroup gap. If group DRO narrows it as much as the crossing penalty, the mechanism story is wrong.
Why not just recalibrate locally at each site, as recurrent local validation recommends? Local recalibration is sound practice and complements this work. It needs labeled local data and a monitoring period, during which the optimistic model is already in use. We will report how many labeled local patients a per-site threshold recalibration of the squared-error model needs before it matches the locked boundary-aware model. That number is the practical value of training-time correction.
Decisive Checks
Three checks carry the proposal, and each can fail.
First, among squared-error models, the site-level and subgroup-level false-reassurance rate should correlate positively with the measured rate of rounding toward normal values. If there is no correlation, the mechanism is wrong even if the loss happens to help.
Second, on locked external evaluation (MIMIC-IV to PhysioNet A and B, and A to B), the crossing loss should reduce the false-reassurance rate relative to squared error, at matched false-alarm rate. It should also beat both the two-sided quantile interval and the symmetric near-boundary ablation. For scale, SafeShift's roughly 10% average collision reduction is the only comparable effect size in the file.
Third, the maximum subgroup gap in false-reassurance rate should shrink more under the crossing loss than under group DRO. This should hold internally on race and insurance and externally on age and sex. The de-rounding ablation should shrink the squared-error model's gap in the same direction.
Following SafeShift's observation of a safety-accuracy trade-off, we will report the MAE cost and downstream sepsis utility alongside every safety gain. A model that removes false reassurance but loses substantial utility will be counted as a failure.
References
- The quality of vital signs measurements in electronic medical records varies by hospital, specialty, and patient demographics
- Neuro-Symbolic World Models for Counter-factual Safety in Automated Driving
- MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV Dataset
- Social-DPF: Socially Acceptable Distribution Prediction of Futures
- VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning
- Trajectory Prediction for Autonomous Driving based on Multi-Head Attention with Joint Agent-Map Representation
- SafeShift: Safety-Informed Distribution Shifts for Robust Trajectory Prediction in Autonomous Driving
- Neural Control System for Continuous Glucose Monitoring and Maintenance
- Early Prediction of Sepsis from Clinical Data: the PhysioNet/Computing in Cardiology Challenge 2019
- Representation Learning for Early Sepsis Prediction
- All models are local: time to replace external validation with recurrent local validation
- An Extensive Data Processing Pipeline for MIMIC-IV
- Early Prediction of Sepsis Using Multi-Feature Fusion Based XGBoost Learning and Bayesian Optimization