The Number That Was Sure

Which statement correctly relates uncertainty, variability and error?

• Variability is what HORIZON should reduce to zero during training; once a large enough dataset removes it, uncertainty and error disappear along with it, since all three ultimately stem from an under-sized sample.
• Uncertainty is incomplete knowledge about a prediction or parameter; variability is the observed differences between individuals that evidences it; error is only measurable once a reference outcome exists, and 11.8 months was presented as though none of the three applied.
• Error is the only one of the three that matters clinically, because uncertainty and variability are statistical abstractions computed over a training population and have no bearing on any individual patient's actual care.
• Variability and error are interchangeable terms for the same phenomenon; the gap between what a model predicted and what happened, and both can, in principle, be measured before a patient's outcome is known.

Select every statement below that correctly identifies a real point where uncertainty enters HORIZON's pipeline.

• The development cohort spans four centres imaged on a single scanner vendor, so the model has never seen data from anywhere it wasn't trained.
• 3,208 patients with incomplete staging were excluded from development, so the model has no grounding in exactly the patients whose work-up is most uncertain.
• Nodule size is read from a radiologist's free-text report rather than remeasured from the image, so the feature itself carries the radiologist's measurement error into the model.
• Missing clinical variables are mean-imputed, which preserves the sample mean but manufactures a value with no evidence behind it for that specific patient.
• The survival head is trained with mean squared error and evaluated on a single 80/20 split with one seed, so both the loss function and the evaluation protocol are themselves modelling choices that were never checked against alternatives.
• Because the mortality head's softmax output sums to one across the two classes, it is mathematically guaranteed to be a well-calibrated estimate of the true probability.
• Once a network finishes training, its weights are fixed, so no further uncertainty can enter between training and the number appearing on the MDT screen.
• Because HORIZON's marketing page states it is '94% accurate,' that published claim is itself evidence that reduces the model's epistemic uncertainty about its own performance.

Two patients, matched on every one of HORIZON's 22 clinical variables and with similar imaging, nonetheless survive for different lengths of time. What kind of uncertainty is this?

• Aleatoric, because health processes are stochastic and no recorded variable determines the outcome.
• Epistemic, because with the right 23rd variable the two patients would resolve into predictably different outcomes.
• Distributional, because it only shows up once the model is deployed outside its training scanner.
• Specification uncertainty, because it depends on the choice to train the survival head with mean squared error.