Melanoma Tumor Size Prediction

Data Information

• Datasets provided by Machine Hack, containing tumor sizes and relevant attributes.

• Training set: 9,146 entries, 10 features

• Test set: 36,584 entries, 10 features

• 0 null values

• Features

  • mass_npea: the mass of the area under study for melanoma tumor.
  • size_npear: the size of the area under study for melanoma tumor.
  • malign_ratio: the ratio of normal to malign surface under study.
  • damage_size: irrecoverable area of skin damaged by the tumor.
  • exposed_area: total area exposed to the tumor.
  • stddevmalign: standard deviation of malign skin measurements.
  • err_malign: error in malign skin measurements.
  • malign_penalty: penalty imposed due to measurement error in the lab.
  • damage_ratio: the ratio of damage to total spread on the skin.
  • tumor_size: size of melanoma tumor.

Exploratory Data Analysis

Distribution of Feature Values

• Apparent correlations most likely due to the inherent proportionality between size and mass.

• Mean of malign_ratio ≈ 0.3, indicating high prevalence of malignancy in the dataset.
• Left-skewed distribution of damage_ratio supports this notion.

Pearson Correlations

• Notable tumor_size correlations

  • size_npear
  • malign_ratio
  • damage_size

Modeling

Overview

• Nature of task

  • Supervised learning
  • Regression to predict numerical value

• Machine learning tools used

  • Scikit-Learn
  • Keras

Procedure

I. Data Preprocessing

1. Training/Validation Split (70%: 30%)
2. Standardization of features with StandardScaler.

II. Hyperparameter Tuning with Randomized Search

• cv = 3
• n_iter = 50
• scoring = ‘neg_mean_squared_error’

III. Training with Tuned Hyperparameters

IV. Performance Evaluation

• Evaluation metric

  • MSE
  • R2

• Models trained and evaluated

  • Multiple Linear Regression
  • Random Forest
  • Support Vector Machine
  • Multi-Layer Perceptron
  • Keras Regression

Model Comparison

• High performance model: Random Forest, Keras Regression

Performance on Test Set

• Best model: Random Forest

Features of Importance

• Primary features of importance

  • malign_ratio
  • damage_size
  • malign_penalty

Assumptions/Limitations

• Assumption

  • All measurements have been obtained with sufficient accuracy/precision.

• Limitations

  • Insufficient number of entries for training set compared to those of test set.
  • Low number of features.

Conclusion

• Best model: Random Forest

• Primary features of importance: malign_ratio, damage_size, malign_penalty

• Prospective improvements

  • Larger dataset
  • Hyperparameter tuning with different techniques
  • Further experiment with Neural Network/Deep Learning models