Work

Final-year project

Retrospective machine-learning stratification of a Brugada-suspect cohort into SCN5A-mediated genetic architecture classes using ICD-10 encoded hospital trajectories

The project used genomics-linked hospital episode material to ask whether historical hospital-coded trajectories could support exploratory stratification in a Brugada-suspect research cohort.

Analytical design diagram for the final-year project.
Analysis design: cohort, hospital-event features, model comparison.

What I was asked

The project asked whether pre-sequencing hospital-coded trajectories could describe useful signal in a Brugada-suspect cohort, and whether that signal was strong enough to support careful retrospective stratification.

What I did

  1. Cohort

    Pre-sequencing hospital episode records

  2. Features

    ICD/OPCS groupings, temporal windows and utilisation features

  3. Observability controls

    Record depth, available history and sparse coding

  4. Split

    Imputation, scaling and encoding were fitted from training data only.

  5. Models

    Logistic regression, random forest and shallow neural-network baselines

  6. Review

    Thresholding, repeated split checks, model metrics, feature-effect outputs and subgroup/fairness review

What it changed

The available hospital record supported a careful exploratory analysis, but not reliable individual classification. The strongest outcome was the workflow itself and the judgement used to recognise when sparse, unevenly observed signal was not strong enough to carry a larger claim.

Testing a second hospital-code representation

After the dissertation, I returned to one narrower question: had the grouped HES representation hidden a small amount of promoter-linked signal that a different encoding could expose? I set up a staged mutual information screen rather than opening an unrestricted feature search.

The first stage tested individual binary hospital-code features with family-wise error control. None passed the declared threshold. The second stage then asked a more practical question: did any complete binary representation improve group-aware prediction beyond a prevalence-only reference? The average log-loss gains were negative, both for HES alone and when the declared nuisance variables were included.

The result did not support progressing to the planned dosage screen. I closed the lane at that point rather than expanding the search until something looked interesting. The useful outcome was a tested negative result.