Machine Learning for Adverse-Event Detection in Latent Tuberculosis Treatment
About a quarter of the world’s population carries a latent tuberculosis infection, which has a 10 to 15 percent chance of progressing to active, contagious disease. Preventive therapy is central to the World Health Organization’s elimination goals, but it can cause adverse events serious enough to make patients abandon treatment. The thesis asked whether the patients most likely to suffer those adverse events could be identified ahead of time from their clinical data, so that therapy could be better targeted.
An earlier study by Campbell and colleagues had modelled this cohort with logistic regression. The hypothesis here was that non-linear machine-learning models, better able to capture complex interactions, might improve on that baseline and turn adverse-event prediction into something clinically actionable.
Working from a clinical dataset of 6,485 treated latent-TB patients, the published logistic regression was first reproduced and then given something it had lacked, an unbiased evaluation protocol separating the data used to fit the model from the data used to judge it; on that footing the baseline reached an area under the curve of 0.65 ± 0.04. Two non-linear models were then trained and compared against it. Both matched the baseline and neither significantly beat it. The reason is visible in the data rather than the models: the cohort is severely imbalanced, with adverse events numbering in the tens against thousands of patients, and those who suffered them do not separate from those who did not in the recorded clinical variables.
The hypothesis was not supported: added model capacity bought nothing, because the limitation lies in the data and not in the linearity of the baseline. The answer to the opening question is accordingly negative — reliable advance warning of adverse events is not achievable from this cohort — and the useful part of that answer is its diagnosis, which tells the designers of the next study what to change: a larger and less imbalanced cohort, described by richer explanatory variables than routine clinical records provide. The reproducible database interface and evaluation package built along the way mean the next attempt starts from a verified baseline rather than a published number.