Continuous Numerical Prediction in Medical AI: Why Tokenisation Is the Wrong Abstraction for Clinical Measurement

When a clinician measures interventricular septal thickness on an echocardiogram, they need a number accurate to within a millimetre. When a vision-language model produces that number by sampling from a probability distribution over text tokens, it is doing something structurally different from what the clinical task requires. This article examines why that structural difference matters, what architectural approaches address it, and what engineering teams need to evaluate before committing to a model architecture for clinical measurement applications.

Companion piece to our broader work on clinical AI validation. See AI Diagnostic Models in Healthcare: Hidden Validity Failures for the evaluation rigour engineering leaders must enforce before deployment.

The Tokenisation Problem in Quantitative Clinical Tasks

Language models generate output by predicting sequences of discrete tokens. A number like "12.4" is not treated as a point on a continuous real-valued axis. It is treated as a string of characters, each selected from a vocabulary of tens of thousands of candidates, with no inherent relationship between "12.4" and "12.5" encoded in the output layer.

This works acceptably when numerical outputs are incidental to the task, for example, when a model is summarising a report and the exact figure is already present in the input. It fails structurally when the model must produce a precise measurement derived from visual evidence, because the loss function during training has no way to express that predicting "13.1" when the ground truth is "12.4" is a smaller error than predicting "9.8".

The commercial implication is that models fine-tuned on clinical measurement tasks using standard autoregressive training will appear to perform well on coarse benchmarks while accumulating systematic errors at the precision level that clinical workflows actually require.

Why Standard Fine-Tuning Does Not Fix the Problem

The reflex response to measurement error in a fine-tuned VLM is to add more training data or adjust the prompt. Neither addresses the underlying architecture. The output layer is still sampling tokens, and the training objective is still cross-entropy over discrete predictions.

Some teams attempt to work around this by binning continuous measurements into categorical ranges and treating the task as classification. This trades one problem for another. Binning introduces artificial discontinuities at bin boundaries and discards the ordinal structure of the measurement space entirely.

Direct regression from a frozen language model backbone using a lightweight MLP head is a more principled approach, but it decouples the regression head from the rich contextual representations that make VLMs useful for prompt-conditioned tasks. The model loses the ability to condition its measurement output on natural language instructions describing which anatomical structure to measure and how.

Hybrid Autoregressive-Diffusion Architectures

The more principled solution is to separate the output modality from the backbone architecture. Autoregressive generation handles language tasks, including prompt interpretation and report generation. A continuous prediction module, operating over real-valued outputs rather than token logits, handles the measurement itself.

HAD-VLM, described by Huang, Guo et al. (arXiv 2026), implements this pattern by pairing an autoregressive language backbone with a diffusion process that predicts a five-component representation: the coordinates of caliper endpoints and the resulting measurement length. The diffusion process operates in continuous space, which means the training signal is sensitive to the magnitude of prediction errors rather than treating all incorrect tokens as equally wrong.

The reported results are instructive. Across four echocardiographic measurement tasks on the EchoNet-LVH dataset, HAD-VLM reduced mean absolute error by 21.4 to 37.6 percent relative to a Vicuna-based autoregressive baseline, and outperformed both fine-tuned Qwen3-VL and direct MLP regression. Critically, medical question answering performance was preserved, which confirms that the hybrid architecture does not require sacrificing language capability to gain measurement precision (Huang, Guo et al., arXiv 2026).

Coordinate-Length Consistency as a Post-Processing Constraint

One underappreciated element of the HAD-VLM approach is the post-processing strategy that enforces geometric consistency between predicted endpoint coordinates and predicted measurement length. In anatomical measurement tasks, these two outputs are not independent. The distance between two predicted caliper positions must be consistent with the predicted scalar measurement.

Enforcing this constraint at inference time acts as a soft anatomical prior. It reduces the variance of individual predictions and provides a natural mechanism for ensemble aggregation across multiple diffusion samples. Teams building similar systems should consider what domain-specific geometric or physiological constraints can be applied at the output stage, because these constraints are often more reliable than additional training data for improving precision at the tail of the error distribution.

Visual Prompt Tuning for Ultrasound Domains

Standard vision encoders in VLMs are pre-trained on natural images. Ultrasound images differ in fundamental ways: they are grayscale, heavily speckled, and encode anatomical information through acoustic impedance gradients rather than colour or texture boundaries. A vision encoder that has not been adapted to this distribution will produce feature representations that are poorly calibrated for echocardiographic measurement tasks.

Visual prompt tuning addresses this by inserting learnable parameters into the vision encoder's input space, adapting its representations to the target imaging modality without retraining the full encoder. This is computationally efficient and preserves the general-purpose representations learned during pre-training while shifting the encoder's attention toward features that are diagnostically relevant in ultrasound.

The practical implication for procurement decisions is that a VLM vendor claiming strong performance on natural image benchmarks is providing limited evidence of ultrasound measurement capability. The vision encoder adaptation strategy is a specific technical question that should be answered before deployment, not inferred from general benchmark scores.

What Engineering Teams Should Evaluate Before Procurement

The architectural questions raised here translate into concrete evaluation criteria. Teams assessing VLMs for clinical measurement tasks should establish whether the model's output layer operates in continuous or discrete space for numerical predictions, and whether the training objective is sensitive to the magnitude of measurement errors.

They should also assess whether the vision encoder has been adapted to the specific imaging modality in question, and whether the model's measurement outputs can be conditioned on natural language prompts describing anatomical targets. A model that can only measure a fixed set of pre-specified structures is not the same product as one that can respond to prompt-specified measurement tasks.

Finally, validation datasets matter more than headline benchmark scores. Error distributions on held-out clinical populations, stratified by image quality and anatomical variation, are the appropriate evidence base for a clinical measurement system. Aggregate mean absolute error on a single benchmark dataset does not capture the tail behaviour that determines whether a system is safe to deploy in a clinical workflow.