Machine learning in clinical pharmacology

Clustering depression diagnoses and predicting SSRI remission

Three R analyses on synthetic cohorts: clustering gene-expression profiles by depression diagnosis, then predicting SSRI remission from clinical, symptom and biological data with four classifiers. Figures are made with matplotlib and seaborn. You can hover, zoom and click the legend on the PCA scatter plots.

Findings

  1. Clusters match diagnosis. Four-cluster k-means, all three hierarchical linkages and the BIC-selected GMM reproduce the four diagnostic groups exactly.
  2. Clinical factors alone. With only sociodemographic factors and MDD severity, the linear SVM gets test AUC 0.90 and accuracy 0.82 (no-information rate 0.50).
  3. Adding biomarkers. Test AUC goes up to 0.95–1.00. At the default 0.5 threshold, though, the RBF SVM calls every patient a non-remitter and gradient boosting finds only about a quarter of remitters.
  4. External cohort. In analysis 3, biomarker levels in the external test cohort are about 30% lower than in training. The models still rank patients in the right order, but the predicted probabilities shift down.

Test ROC AUC across all models

Held-out test set, with and without biomarkers. The dashed line is chance.

Grouped bar chart of test ROC AUC for four models in analyses 2 and 3

The three analyses

  1. Gene-expression clustering. 30 subjects × 10 genes. PCA, k-means, hierarchical clustering and Gaussian mixture models, scored by how well the clusters recover the clinical diagnosis (adjusted Rand index).
  2. SSRI remission. 400 patients with major depressive disorder, random 75/25 split. Random forest, linear and RBF SVMs and gradient boosting, trained on clinical factors with and without 9 biological markers.
  3. Symptom items. The same question with 12 baseline QIDS-SR symptom items added, a predefined 300-patient training cohort and a 100-patient external test cohort.

Evaluation

  • Hyperparameters were tuned with 5× repeated, stratified 10-fold cross-validation on the training set only.
  • Each model was scored once on the held-out test set.
  • Remission (response = 1) is the positive class for sensitivity, specificity and PPV/NPV.
  • Test accuracy is compared with the no-information rate (the majority-class share) using a one-sided exact binomial test.

The data are synthetic teaching datasets and aren't redistributed here. Only derived results are published.