🧠 Building an AI-Powered Disease Classifier: A Machine Learning Journey

🩺 Introduction

In today’s healthcare world, early and accurate diagnosis can be life-saving. But symptoms often overlap — fever could signal the flu, COVID-19, or pneumonia. Even seasoned clinicians can struggle to differentiate them.

So I asked: Can machine learning help?

To find out, I built an end-to-end AI system that predicts diseases based on symptoms and risk factors. Along the way, I compared K-Nearest Neighbors (KNN) and Logistic Regression.

In this blog, you’ll discover:

✅ Data preprocessing & feature engineering ✅ Model training & evaluation ✅ Key insights & deployment ✅ Real-world applications

Let’s dive in! 👇

1️⃣ The Problem & Approach

🎯 Objective

Build a machine learning model that:

• Accepts symptoms (e.g., fever, cough) and risk factors (e.g., diabetes, smoking)
• Predicts the most likely disease
• Compares the performance of KNN vs. Logistic Regression

📊 Dataset

The dataset included:

• Symptoms: Text-based (e.g., “nausea, dizziness”)
• Risk Factors: Categorical (e.g., hypertension: yes/no)
• Diseases: Labeled outcomes (e.g., COPD, Diabetes)

2️⃣ Data Preprocessing & Feature Engineering

🧾 Text Features: Symptoms

Symptoms were vectorized using TF-IDF (Term Frequency-Inverse Document Frequency), which:

• Converts raw text into numerical format
• Gives higher weight to rare (but significant) terms like “hemoptysis”

from sklearn.feature_extraction.text import TfidfVectorizer

tfidf = TfidfVectorizer(token_pattern=r'\b\w+\b', stop_words='english')
symptoms_tfidf = tfidf.fit_transform(df['Symptoms'].apply(' '.join))

🧬 Categorical Features: Risk Factors

Risk factors were encoded using One-Hot Encoding:

from sklearn.preprocessing import OneHotEncoder

encoder = OneHotEncoder()
risk_factors_encoded = encoder.fit_transform(df[['Hypertension', 'Diabetes']])

🧩 Final Feature Set

We combined text and categorical features using hstack:

from scipy.sparse import hstack

X = hstack([symptoms_tfidf, risk_factors_encoded])
y = df['Disease']

3️⃣ Model Training & Evaluation

🧪 Algorithms Tested

Model Hyperparameters Tested KNN k = 3, 5, 7; Metrics = Euclidean, Manhattan, Cosine Logistic Regression Penalty = L1/L2, class_weight balancing

⚙️ Training Process

• 5-fold cross-validation for robustness
• Stratified sampling to address class imbalance
• Metrics: Accuracy, Precision, Recall, F1-score

from sklearn.model_selection import cross_validate
from sklearn.neighbors import KNeighborsClassifier

knn = KNeighborsClassifier(n_neighbors=3, metric='cosine')
cv_scores = cross_validate(knn, X, y, cv=5, scoring=['accuracy', 'f1_weighted'])

📈 Results

Model Accuracy F1-Score Best Configuration KNN 87% 0.85 k=3, Cosine similarity Logistic Regression 83% 0.81 L2 penalty, class_weight='balanced'

💡 Why KNN won: Cosine similarity handled symptom vectors well, while Logistic Regression struggled with rare diseases.

4️⃣ Deployment & Real-World Use Cases

💾 Saving the Model

Used joblib to serialize the trained model:

import joblib

joblib.dump(knn, 'disease_predictor.pkl')

🔍 Prediction Example

from predictor import DiseasePredictor

model = DiseasePredictor.load('disease_predictor.pkl')
model.predict([["fever", "cough", "fatigue"]])  # ➜ ['Influenza']

🌍 Real-World Applications

1. Telemedicine Platforms: Pre-consultation triage
2. Hospital Support Tools: Assist doctors with differential diagnosis
3. Public Health Systems: Track symptom clusters and disease spread

5️⃣ Challenges & Lessons Learned

✔️ Class Imbalance: Handled using stratified cross-validation ✔️ Text Representation: TF-IDF outperformed Bag-of-Words ✔️ Interpretability vs. Accuracy: Logistic Regression was easier to explain, but KNN performed better

🧪 What’s Next?

• Experiment with BERT embeddings for smarter text understanding
• Integrate lab test results and vitals for improved accuracy

✅ Conclusion

This project showed that machine learning can assist clinicians by suggesting potential diagnoses. While not a replacement for doctors, it can support faster, smarter decisions.

🛠️ Final Result:

• Best model: KNN (Cosine Similarity + TF-IDF)
• Use cases: Telehealth, diagnostics, public health monitoring

👉 Try it yourself: GitHub Repo