Skip to main content

Skills application solution

Instruction and application
In Progress

Skills application solution​

Compare your skills application output to the solution example below provided by Multiverse subject matter experts.

Solution In this solution, you applied a robust and realistic data handling workflow using the MediCareAI dataset.1. Data splitting and leakage prevention

  • Used a temporal holdout split, training on data before July 2024 and testing on the following six months.

  • Verified that no admission dates in the test set precede those in the training set, confirming no temporal leakage.

  • Applied stratified 5-fold cross-validation within the training set for hyperparameter tuning to maintain balance across folds.2. Handling class imbalance

  • Used SMOTE (Synthetic Minority Oversampling Technique) to generate synthetic examples of readmitted cases.

  • Verified post-resampling balance with approximately 50:50 class distribution in the training set.

  • Ensured that SMOTE was applied only on the training data to avoid leakage into the test set.3. Data quality protocols

  • Performed median imputation for missing numerical values (BMI, glucose, blood pressure).

  • Capped extreme glucose and blood pressure values at the 1st and 99th percentiles.

  • Removed 80 duplicate rows to eliminate redundant records.

  • Validated final data integrity by checking distributions and summary statistics after cleaning.

Outcome: A clean, balanced, and temporally consistent dataset was prepared for model training. The solution improves model generalisation and fairness while maintaining realistic conditions for evaluation.

Why this solution works well​

  • Demonstrates methodical leakage prevention by using a time-based split.
  • Balances classes effectively using SMOTE without distorting distributions.
  • Implements reproducible data quality checks that could easily be automated in production.

Tips for applying this skill in your role​

  • Always design data splits that mirror real-world prediction scenarios (e.g., future vs. past data).
  • Handle imbalance carefully, oversampling and weighting are powerful but must be isolated to training data.
  • Build modular preprocessing pipelines that include data validation checks before training or deployment.

Action item: Reflection​

  • Compare your output to the solution example provided. What did you do well? Where could you improve?
  • How might you automate or scale your data quality and leakage prevention steps across multiple ML pipelines in your organisation?
In progress