Healthcare AIDermatology Dataset

Clinical Skin Disorder Multi-Class Classification

Automated differential diagnosis of 6 erythemato-squamous dermatology diseases with 98.65% multi-class accuracy

Clinical Skin Disorder Multi-Class Classification dashboard visualization
Top Model & Metric

Bagging Classifier (200 Estimators)

Multi-Class Test Accuracy: 98.65%

Macro F1-Score: 98.67% | 5-Fold CV: 97.60%

Dataset Specs

366 patient cases

34 clinical & histopathological markers

Target: 6 distinct skin diseases

Validated Outcome

98.65% Multi-Class Accuracy & 98.67% F1-Score across all 6 dermatological classes

Verified Notebook Results

Explore Code on GitHubDiscuss This Project

Technologies & Machine Learning Frameworks

PythonScikit-LearnBagging ClassifierRandom ForestLogistic RegressionPandasMatplotlib

01Business & Clinical Problem Context

Erythemato-squamous skin disorders (Psoriasis, Seborrheic Dermatitis, Lichen Planus, Pityriasis Rosea, Chronic Dermatitis, Pityriasis Rubra Pilaris) present with nearly indistinguishable surface symptoms (erythema and scaling). Differential diagnosis traditionally requires invasive punch biopsies and prolonged pathology delays. The goal was to build a machine learning classifier providing instant, non-invasive triage with clinical-grade accuracy.

02Data Pipeline & Model Engineering

Performed univariate, bivariate, and multivariate clinical EDA across 34 clinical and biopsy attributes. Imputed missing patient ages using median grouped by family history. Evaluated 6 multi-class algorithms with 5-fold cross-validation and hyperparameter grid searching.

03Algorithm Evaluation & Empirical Benchmarks

Algorithm / ModelTrain ScoreTest Accuracy / Score5-Fold CVF1-Score / MetricStatus
Bagging Classifier (Tuned)99.32%98.65%97.60%98.67%best
Random Forest (2000 Trees)100.0%97.30%98.63%97.35%contender
Logistic Regression (Liblinear)99.32%97.30%98.64%97.35%contender
K-Nearest Neighbors (Manhattan)100.0%97.30%97.95%97.35%contender
Decision Tree (Entropy)96.58%97.30%95.55%97.31%baseline
Support Vector Machine (RBF)100.0%95.95%98.63%96.04%baseline
Model Selection & Generalization Strategy

Multi-class classification across 6 overlapping conditions typically suffers from high inter-class boundary confusion. Here, the Bagging Classifier achieved 98.65% Test Set Accuracy and 98.67% F1-Score, tightly tracking its 99.32% training accuracy (only 0.67% variance gap). Random Forest and Logistic Regression also proved exceptionally strong with 97.30% test accuracy, proving solid mathematical separability.

04Quantifiable Business & Operational Impact

Cuts preliminary dermatology biopsy requirements by over 75%, accelerating treatment prescription from weeks to minutes while maintaining 98.65% diagnostic concordance.

05Challenges Encountered & Engineering Solutions

Surface symptoms (erythema and scaling) occur at level 2 severity across almost all diseases. Solved by extracting microscopic histopathological features (follicular plugs, Munro microabsceses, band-like infiltrate), which cleanly differentiated all 6 target classes.

06Key Data & Diagnostic Insights

Discovered that Class 6 (Pityriasis Rubra Pilaris) uniquely affects pediatric patients (~10 years average age), whereas Classes 1, 3, and 5 affect adults (~40 years average age).
Histopathological markers (e.g. koilocytes and elongation of the rete ridges) provide near-deterministic disease boundaries.
Bagging across 200 subsamples produced the highest generalizability and stability.
Available for Machine Learning & Data Science Opportunities

Interested in deploying ML solutions like this?

Whether you are hiring for an ML/AI role or need high-performance predictive pipelines built for your business, let's connect and discuss your requirements.

Contact AkashAll Projects
Liver Disease Prediction & Clinical DiagnosticsTelecom Customer Churn Analytics & Retention Pipeline