Foundations capstone

The first set of questions concerns the training data. Drawing on Modules 1 and 2 (what data is, data representation):

1. Population match. Was the model trained on a population similar to the trust's patients? A model trained on a tertiary referral centre (sicker patients, more complex conditions) may not transfer to a district general hospital. A model trained in the US may not generalise to UK populations with different demographics, comorbidity profiles, and healthcare pathways.
2. Feature availability. Does the trust collect the same vital signs and lab results at the same frequency as the training data? If the model expects continuous pulse oximetry but the trust takes spot readings every 4 hours, the input distribution will differ from training.
3. Label quality. How was "deterioration" defined in the training data? Was it transfer to ICU, activation of a rapid response team, or death within 24 hours? Different definitions produce different models with different performance characteristics.
4. Missing data. Hospital data is notoriously incomplete. How did the vendor handle missing vitals? Imputation methods matter: mean imputation can mask the very signals the model needs to detect.

The vendor reports AUC 0.89 and F1 0.78. Drawing on Module 5 (evaluating AI):

1. AUC is a threshold-independent metric. It summarises performance across all possible decision thresholds. But the trust needs to operate at a specific threshold: how sensitive should the system be? What false positive rate is clinically acceptable? AUC does not answer these questions. You need the full precision-recall curve and the receiver operating characteristic (ROC) curve, not just a single number.
2. Class imbalance. Deterioration events are rare (typically 2-5% of hospital admissions). On imbalanced data, high AUC can coexist with clinically unacceptable precision. If the model fires 100 alerts per day and only 5 are genuine, staff will stop responding. You need the positive predictive value (PPV) at the operating threshold.
3. Temporal validation. Was the model validated on data from the same time period as training (random split) or from a later period (temporal split)? Random splits inflate performance because they leak temporal patterns. Temporal validation is the only acceptable approach for time-series clinical data.
4. Cross-validation. Was performance reported from a single split or from k-fold cross-validation with confidence intervals? A single split gives you a point estimate with no measure of reliability.

The reason a single headline number cannot settle the threshold is that the threshold produces two separate harms, and they land on different patients. The figure below follows both from the moment the score is applied on the ward: an alert on a stable patient, and silence on a deteriorating one. Each path ends at a named harm and at the control that answers that harm alone.

Drawing on Module 6 (deep learning architectures):

1. Temporal data requires temporal architectures. Vital signs are time series. An architecture that treats each measurement independently (e.g. a simple feedforward network on the latest vital signs) discards temporal trends that are often the earliest signal of deterioration. LSTMs, temporal convolutional networks (TCNs), or transformers are more appropriate because they model how vital signs evolve over time.
2. Irregularly sampled data. Hospital vital signs are not measured at fixed intervals. A patient in a general ward might have observations every 4 hours; a patient causing concern might have them every 30 minutes. The architecture must handle irregular time intervals, either through explicit time-aware mechanisms or through appropriate preprocessing.
3. Interpretability versus performance. A gradient-boosted tree model may sacrifice 1-2 points of AUC compared to a deep neural network but produce feature importance rankings that clinicians can understand and trust. In healthcare, interpretability is not optional: clinicians need to know why the system flagged a patient to decide whether to act.

Drawing on Module 7 (responsible AI basics):

1. Disaggregated performance. Does the model perform equally well across age groups, ethnicities, and sexes? Elderly patients, who are most likely to deteriorate, may also be the group most poorly represented in training data. A model that performs well on average but poorly for patients over 80 is actively dangerous.
2. Alert equity. If the model's false positive rate differs by demographic group, some patients receive unnecessary interventions while others receive insufficient attention. This is the equalized odds problem from Module 7 applied to clinical care.
3. Model card. Has the vendor published a model card? Does it include disaggregated metrics, intended use constraints, and known limitations? If not, the trust is deploying a system without the minimum documentation needed for accountability.
4. Override policy. Can clinicians override the model? What happens when they do? An AI system that generates alerts but provides no mechanism for clinical judgment to prevail is unsafe. Conversely, if clinicians override every alert, the system adds cost without value.