Built an end-to-end healthcare analytics pipeline, from clinical data cleaning to machine learning and an interactive R Shiny dashboard, using real patient data from Pakistan.

Existing heart disease prediction models are almost universally built on Western datasets like Cleveland, which don't reflect Pakistani patient demographics or comorbidity patterns. I sourced a real clinical dataset of heart patients in Pakistan (several hundred records, 60+ clinical variables) and built the full pipeline myself: cleaning, statistical testing, predictive modeling, patient risk segmentation, and an interactive R Shiny dashboard that turns raw clinical data into a clinical decision support tool, not just a report.

Standardized inconsistent free-text symptom fields (e.g. "chest pain, sneezing and cough" vs "sneezing, chest pain and cough") by tokenizing, alphabetizing, and rewriting them in a consistent order, so identical symptom combinations weren't fragmented into false categories

Validated and encoded categorical/binary variables, checked numeric fields against physiological limits, and confirmed zero missing values before modeling

Ran a Shapiro-Wilk normality test across all variables first, letting the actual distribution of the data (not assumption) decide which tests to use downstream

Applied Chi-Square tests for categorical associations (symptom type vs severity, lifestyle factors vs severity), Kruskal-Wallis and Mann-Whitney U for non-normal group comparisons (age vs severity, blood markers by gender), and Spearman correlation for non-linear relationships between biomarkers

Built a Multinomial Logistic Regression and a Random Forest model to predict disease severity, pairing an interpretable baseline with a model that captures non-linear feature interactions and ranks feature importance

Built a Binary Logistic Regression model to predict mortality, using univariate screening (p < 0.1), VIF-based multicollinearity checks, and stepwise selection (stepAIC) to arrive at a lean, interpretable final model

Converted the mortality model's coefficients into a single weighted patient risk score, giving clinicians one number rather than a black-box probability

Applied K-Means clustering on key clinical markers (age, serum creatinine, CK-MB, cholesterol, resting BP, max heart rate) to segment patients into Low, Moderate, and High-Risk groups

Used PCA to reduce dimensionality and visualize cluster separation in a 2D interactive scatter plot

Built a 4-module interactive R Shiny dashboard (EDA, Hypothesis Testing, Models, Policy Suggestions) with dynamic tooltips and interactive plots, translating every statistical and ML output into a form a clinician or policymaker, not just a data scientist, can explore and act on

Skills R Programming · Data Cleaning & Standardization · Exploratory Data Analysis · Statistical Hypothesis Testing (Chi-Square, Kruskal-Wallis, Mann-Whitney U, Spearman Correlation) · Multinomial & Binary Logistic Regression · Random Forest · Feature Engineering · Multicollinearity Analysis (VIF) · K-Means Clustering · Principal Component Analysis (PCA) · Interactive Dashboard Development · Healthcare Analytics · Data Storytelling

Tools R · Shiny · shinydashboard · dplyr · tidyr · ggplot2 · plotly · randomForest · caret · MASS · nnet · cluster