Machine Learning · Healthcare
Skin Cancer Diagnosis Classifier with Decision Trees
A supervised learning project using Decision Tree and Random Forest classifiers to distinguish between various skin disorders based on clinical features.
Overview
This project uses a dermatology dataset to build classification models for predicting a disorder type of the erythemato-squamous skin disease. The focus is on understanding the full ML pipeline: preprocessing and exploratory data analysis, model training, performance evaluation and analyzing feature importance.
This demonstrates how classification models can support decision-making within the healthcare field and help professionals diagnose certain diseases.
Data & Methods
Dataset
- Dataset containing clinical features of various skin lesions.
- Target label indicating 1 of 6 possible disorder diagnoses.
Preprocessing
- Loaded CSV into pandas and performed exploratory data analysis, checked missing values, and distributions
- Dropped missing values or features as needed. Age and Family History features were dropped to focus on physical traits.
- Examined target feature correlation before training to get an idea of potential feature importance.
Modeling
- Split data into train and test sets and created dummy variables.
- Trained a default DecisionTreeClassifier using scikit-learn.
- Tuned hyperparameters using a GridSearchCV to find optimal values and trained a DecisionTreeClassifier using these values.
- For comparision, I tuned hyperparameters for a Random Forest model as well.
- Evaluated all models using accuracy, precision, recall, and confusion matrix.
Tech Stack
Python, scikit-learn, pandas, NumPy, matplotlib, Jupyter
Metrics
Default parameters DecisionTree Classifier:
- Accuracy: 93%
- Precision (weighted average): 94%
- Recall (weighted average): 93%
Tuned parameters DecisionTree Classifier:
- Accuracy: 93%
- Precision (weighted average): 94%
- Recall (malignant class): 93%
Tuned parameters DecisionTree Classifier:
- Accuracy: 96%
- Precision (weighted average): 97%
- Recall (weighted average): 96%
Key Charts
Export these plots from the notebook and save them in assets/ with matching filenames.
Analysis
-
From these results we can see that the tuned Random Forest Classifier performed better in total Accuracy,
Precision, and Recall. There was no difference between the Decision Tree classifiers as seen in the confusion matrices.
Challenges & Learnings
- Balancing model complexity vs interpretability in a healthcare-related problem.
- Understanding how class imbalance affects metrics like accuracy vs recall.
- Using visual tools like confusion matrices to explain misclassifications.