AutoML for clinical research

Go to the Quick Start ↗
AUTOML FOR CLINICAL RESEARCH

Discover what IAML brings to your research.

A complete study with a few lines of Python.
Train, evaluate and explain with the built-in pipeline.

Scroll to explore

You bring the question and the data.

Can these measurements distinguish malignant from benign samples?
Scroll to follow IAML from this table to an explained prediction.

STUDY TABLE120 observations · 6 features
RadiusTextureOutcome
11.2514.78Benign
14.5427.54Malignant
12.4624.04Malignant

Breast cancer dataset bundled with scikit-learn.
Here, 1 = malignant and 0 = benign.

SET ASIDE THE TEST DATA
90 for training30 held out
IAML(...).fit(X_train, y_train) Keep scrolling

IAML searches. Compares. Fits.

The default pipeline searches preprocessing and predictors together. IAML compares candidates by cross-validation, then fits the selected pipeline.

Training data90 observations IAML Fitted pipelineReady to predict
CANDIDATE

XGBoost

0.990Cross-validation ROC AUC
SELECTED

CatBoost

0.995Cross-validation ROC AUC

The results become visible.

One call generates the performance figures. These use the 30 observations kept outside the search.

0.967 test ROC AUC29 / 30 correct labels0.909 sensitivity

Balanced accuracy 0.955F1 0.952Precision 1.000Specificity 1.000Balanced error 0.045

ROC curve View larger ↗ROC curve: AUC 0.967 on the held-out data. Confusion matrix View larger ↗19 benign and 10 malignant observations correctly classified, with one malignancy missed. Precision–recall View larger ↗Precision–recall curve for malignancy, with average precision rounded to 0.96.
BIBLIOGRAPHY EXAMPLE
  1. Prokhorenkova et al. (2017). CatBoost: unbiased boosting with categorical features. arXiv:1706.09516.
  2. Lundberg & Lee (2017). A Unified Approach to Interpreting Model Predictions. arXiv:1705.07874.

And the predictions become understandable.

IAML computes SHAP contributions for the fitted pipeline: inspect one prediction, then compare contributions across observations.

Inside one prediction View larger ↗Six SHAP contributions lower one observation's predicted probability of malignancy from 37.8% to 0.1%.
FOR THIS OBSERVATION37.8% → 0.1%

From the background average to the predicted probability of malignancy. The blue contributions lower it.

Across five observations View larger ↗Mean absolute SHAP contributions across five held-out observations.

Try the complete workflow.

A fitted pipeline, performance metrics and plots, and feature explanations — connected in one Python workflow.

  • Your trained pipeline
  • Evaluation metrics
  • Model performance plots
  • SHAP explanations of predictions
  • Method bibliography

IAML supports classification, regression and survival analysis.

Run the Quick Start → See the example and code

Optional customizationGo further: build your own pipeline →

Scroll to follow the analysis ↓Precalculated example · Methods and context ↗