Point-of-care cardiac ultrasound is increasingly run by non-specialists, and a natural aid is an agent that decides, clip by clip, whether it has seen enough views to make a diagnosis or should ask for another. The only published agent of this kind, for aortic stenosis, maximizes average reward. Averages are exactly what break when a model moves from a curated archive to a live ward. Retrospective cohorts over-represent clean, canonical cases. Live bedside scanning produces more ambiguous valves, more foreshortened views and more probe jitter. An agent that stops early because "on average one clip is enough" will fail on precisely the cases that become more common after deployment.
This note proposes to test one specific claim. Training the stopping agent on conditional value-at-risk (CVaR, the average outcome over the worst α-fraction of patients) rather than expected outcome makes its performance provably insensitive to a specific, measurable kind of case-mix shift. That shift is any prospective population that upweights some subgroup of retrospective patients by at most a factor 1/α. We would build the agent on a lightweight factorized spatiotemporal video encoder so it can run on a portable scanner. We would then check the claim in a silent prospective deployment. In that setting the full view protocol is still acquired, so every stopping decision can be scored against real, non-curated patients without anyone's care depending on the agent.
Core Claim
Let $R$ be a patient's outcome under a stopping policy. It equals minus a per-clip acquisition cost times the number of clips, minus a misdiagnosis loss that is large when significant stenosis is called normal. CVaR has a standard dual form:
$$\mathrm{CVaR}_\alpha(R) = \min_{Q:\; dQ/dP \le 1/\alpha} \mathbb{E}_Q[R],$$
where $P$ is the retrospective patient distribution and $Q$ ranges over every reweighting of it whose density ratio never exceeds $1/\alpha$. A policy that maximizes $\mathrm{CVaR}_\alpha$ on retrospective data therefore maximizes a guaranteed floor on expected outcome under every such prospective population.
The retrospective-prospective gap then becomes a quantity we can estimate: the density ratio between the prospective cohort and the retrospective cohort. This yields a falsifiable prediction. When the estimated ratio's effective bound is below $1/\alpha$, the CVaR policy's prospective expected outcome should not fall below its retrospective CVaR. A risk-neutral policy has no such floor. This mechanism also lets the note distinguish itself from the obvious combination of the three directions. We are not simply "adding CVaR to an echo agent." The claim is that the risk parameter α can be chosen from a measured shift magnitude, and that the resulting robustness can be checked prospectively.
Where This Sits in Prior Work
The motivating work is PRECISE-AS. It casts aortic stenosis diagnosis as active feature acquisition: a Double DQN decides whether to request another parasternal clip or submit to a transformer classifier. It reports 80.6% balanced accuracy using 1.88 of four clips on average. We keep its problem formulation and its missing-view classifier, but replace its risk-neutral objective. Its evaluation was retrospective, and nothing in its objective prevents early stopping on atypical severe cases. A second motivating paper is Straight to the point. Its DQN probe-guidance agent cut false "stop here" commands from 25.9% (supervised baseline) to 1.6%. This shows that stopping errors are the dominant failure of naive approaches in ultrasound. It too works on a static, retrospective grid simulation.
The risk-sensitive method comes from two sources. Predictive CVaR Q-learning defines recursive tail-value and tail-probability functions that give CVaR a Bellman update, together with randomized initial risk budgets. It reaches CVaR 2.45 against a 2.50 optimum on a synthetic decision tree, but was only tested on tabular problems. Our horizon is at most four decisions over precomputed clip embeddings, which is close to the regime where that method was shown to work; we only replace the tables with small networks. We deliberately do not follow WCSAC, which approximates the cost distribution as a Gaussian. Here the outcome is a mixture of a small acquisition cost and a rare large misdiagnosis loss, which is strongly bimodal. The heavy-tail argument of Jung et al. warns that Gaussian CVaR would underestimate the tail in this setting, so we use sample-based quantile estimates. Pessimism Meets Risk constrains the design. Its guarantees for offline risk-sensitive learning require coverage of failure regions. Our setting satisfies that requirement unusually well: every archived study contains all views, so every stopping sequence can be replayed exactly and no extrapolation is needed.
The encoder comes from the factorized-convolution line. A factorized, or (2+1)D, convolution splits each 3×3×3 kernel into a 3×3 spatial filter followed by a 3-frame temporal filter. SpaceTime-SonoNet shows that this matches a monolithic 3D network on ultrasound plane classification (63.54% vs 63.58%) at roughly half the parameters and compute. Hughes et al. use a (2+1)D ResNet on echo and show, by blinding the model to motion, that some targets depend on chamber motion rather than texture. For aortic stenosis, leaflet motion is the signal. The encoder's role is therefore a constraint rather than the contribution: it must be cheap enough for a portable device and must see motion.
Our evaluation design follows the prospective-validation literature. Ng et al. is the template for phased rollout, where the AI acts first as a non-binding extra reader. Chao et al. and Horng et al. argue that offline accuracy does not predict live behavior. The spectrum-bias account of Wang et al. supplies the mechanism we try to hedge against. There is also a directly relevant warning about hardware shift. Foundation versus Domain-Specific Models reports that a domain-specific echo model falls from DSC 0.915 to 0.667 on handheld POCUS, while fine-tuned SAM holds 0.857. That shift is partly a representation failure, which CVaR over outcomes cannot repair. We therefore treat encoder robustness as a separate axis to measure.
Research Program
The work has three parts.
First, we build the agent. A (2+1)D encoder, pretrained on EchoNet-Dynamic, embeds each parasternal long- and short-axis clip. A PRECISE-AS-style transformer classifier maps any subset of embeddings to Normal, Early or Significant stenosis. A CVaR stopping agent is trained offline by replaying every view ordering of every archived study. We train one agent per α in {1.0, 0.5, 0.2, 0.1, 0.05}; α = 1 is exactly risk-neutral, so this sweep doubles as the key ablation.
Second, we measure the shift. We collect a prospective cohort in silent mode at the same institution, with at least one site using handheld POCUS devices. For this cohort we fit a classifier that discriminates retrospective from prospective studies, using the frozen embeddings and acquisition metadata. From it we estimate the density ratio, reporting its 95th and 99th percentiles as the effective bound.
Third, we test the prediction. For each α we compare the agent's retrospective CVaR with its prospective mean outcome. We also decompose the shift into two parts: change in class balance, and change within class (image quality, borderline valves). The theory says CVaR should matter mainly for the second part.
A small human phase follows only if silent mode is favorable. In it, novice operators scan with the agent's "one more view" prompts versus a fixed protocol, and we record exam time and adjudicated misses, in the manner of the reader-study design of Meng et al..
Baselines
The naive reference point is the fixed full protocol: acquire all four parasternal clips and classify. It sets the accuracy ceiling for this classifier and the maximum acquisition cost. We also run a fixed two-clip protocol, one long-axis and one short-axis clip, which matches the average clip budget that learned agents reach.
The strongest existing method is the risk-neutral Double DQN agent of PRECISE-AS, which reports 80.6% balanced accuracy at 1.88 videos per study on its private tertiary dataset. That dataset is not available to us, so we reimplement the agent and run it on our data. It gets the same encoder, the same classifier, the same replay environment and the same hyperparameter search budget as the CVaR agent. Beating it means two things at matched mean clip usage (within 0.1 clips). The first is a lower rate of significant stenosis called normal, together with higher CVaR at α = 0.1, on the prospective silent cohort. The second is that balanced accuracy does not drop by more than one point.
A cost-sensitive variant of the same risk-neutral agent is the most dangerous competitor. It raises the penalty for missed significant stenosis until its retrospective miss rate equals the CVaR agent's. Comparing at matched retrospective misses isolates whether CVaR buys anything prospectively beyond reweighting classes.
A calibrated confidence-threshold stopper is also included. It stops when the classifier's risk estimate passes a threshold set by conformal risk control, in the style of Conformal Risk-Averse Decision Making with OCE Risk Control. It uses no RL. Its threshold is calibrated on the same retrospective data.
Encoder ablations swap the (2+1)D encoder for a single-frame 2D encoder, a monolithic 3D encoder and a fine-tuned foundation encoder, all under the same CVaR agent. This separates representation shift from decision-rule shift.
Why not just…?
Why not just give the risk-neutral agent a much larger penalty for missing severe stenosis? This is the most serious threat, and we cannot rule it out in advance. If the prospective shift is mostly a change in how many patients have significant stenosis, class-weighted rewards should hedge just as well. The dual form above says CVaR's advantage comes from protecting every high-loss subgroup, including hard-to-image severe cases, rather than a fixed class. The cost-sensitive baseline at matched retrospective miss rate, together with the class-balance versus within-class shift decomposition, settles this question. If the two agents tie prospectively when the shift is mostly within class, the CVaR claim fails and we will say so.
Why not just stop when a calibrated classifier is confident enough? Conformal thresholds give finite-sample guarantees only when calibration and deployment data are exchangeable, and the retrospective-prospective gap is exactly a failure of exchangeability. The supervised stopping rule in Straight to the point produced 25.9% false stops, though that was an uncalibrated classifier, not a conformal one. Calibration may therefore close much of that gap. The conformal baseline is run on the prospective cohort precisely to measure how far its guarantee degrades compared with the CVaR floor.
Why not just always acquire all four views? On a cardiology service this is often acceptable. At the point of care, though, each additional parasternal view costs a novice operator time and success is not guaranteed. PRECISE-AS shows roughly half the clips suffice on average. The full protocol stays in the study as the accuracy ceiling, and the human phase measures the time actually saved.
Why not just use a more robust foundation encoder and keep the risk-neutral agent? The POCUS collapse reported in Foundation versus Domain-Specific Models suggests the encoder may be the larger problem. CVaR over outcomes does not fix embeddings that are wrong. The encoder ablation, crossed with α, shows which factor dominates prospective degradation. If a foundation encoder alone closes the gap, the lesson is to invest in representation, not risk-sensitive control.
Proposed Experiments
The retrospective stage uses our institution's archive of parasternal studies labeled for stenosis severity, split by date so that validation data comes after training data. It reports three things for each agent: balanced accuracy, significant-stenosis-to-normal miss rate, and mean clips per study. It also reports CVaR at α = 0.1 of per-patient outcome. For context only, we also report encoder accuracy on EchoNet-Dynamic ejection-fraction regression, alongside the published E-ViM³ result of MAE 3.54 E-ViM³.
The decisive check comes in the silent prospective stage, where every agent is scored counterfactually on the fully acquired studies. The test is whether the CVaR agents' prospective mean outcome stays above their retrospective CVaR whenever the measured density-ratio bound is below $1/\alpha$. A second test is whether the cost-sensitive risk-neutral agent violates the same floor, specifically on the within-class component of the shift. A third test is whether the (2+1)D encoder matches the monolithic 3D encoder's prospective accuracy at about half the compute, as it did retrospectively in SpaceTime-SonoNet.
The prediction is falsified in either of two cases. The first is if the floor is violated despite a small measured ratio. The second is if class-weighted risk-neutral agents match CVaR agents on the prospective tail. Either result is informative. The first would show that the retrospective-prospective gap in echo is not a reweighting of retrospective cases at all, but genuinely new cases that no retrospective objective can anticipate.
References
- PRECISE-AS: Personalized Reinforcement Learning for Efficient Point-of-Care Echocardiography in Aortic Stenosis Diagnosis
- Straight to the Point: Reinforcement Learning for User Guidance in Ultrasound
- Predictive CVaR Q-learning
- WCSAC: Worst-Case Soft Actor Critic for Safety-Constrained Reinforcement Learning
- Quantile Constrained Reinforcement Learning: A Reinforcement Learning Framework Constraining Outage Probability
- Pessimism Meets Risk: Risk-Sensitive Offline Reinforcement Learning
- SpaceTime-SonoNet: efficient classification of ultra-sound video sequences
- Deep Learning Prediction of Biomarkers from Echocardiogram Videos
- Prospective implementation of AI-assisted screen reading to improve early detection of breast cancer
- Evaluating Translational AI: A Two-Way Moving Target Problem
- Creating an automated trigger for sepsis clinical decision support at emergency department triage using machine learning
- Automated segmentation and diagnosis of pneumothorax on chest X-rays with fully convolutional multi-scale ScSE-DenseNet: a retrospective study
- Foundation versus Domain-Specific Models for Left Ventricular Segmentation on Cardiac Ultrasound
- AI support for accurate and fast radiological diagnosis of COVID-19: an international multicenter, multivendor CT study
- Conformal Risk-Averse Decision Making with Optimized Certainty Equivalent Risk Control
- E-ViM 3 : Mamba-3D as Masked Autoencoders for Accurate and Data-Efficient Analysis of Medical Ultrasound Videos