Asif Rahman - Machine / Deep Learning Projects

Project Name: Changes in Psychiatric Diagnosis Associated With SARS-CoV-2 Infection and Predicting the Development of New Psychiatric Illness in COVID Patients by Using Machine Learning Approach: A Study Using the US National COVID Cohort Collaborative (N3C)

NIH Project

Summary:

This project investigates the complex link between SARS-CoV-2 infection and the onset of new psychiatric disorders, with a focus on Schizophrenia Spectrum and Psychotic Disorders (SSPD). Using data from the U.S. National COVID Cohort Collaborative (N3C), I developed predictive models to identify high-risk COVID-19 patients who may be prone to psychiatric complications post-infection. By comparing COVID-19 patients with Acute Respiratory Distress Syndrome (ARDS) and COVID-negative control groups, I established a significant association between COVID-19 and psychiatric outcomes, reinforcing the need for proactive mental health interventions in COVID-19 recovery. My analysis, utilizing machine learning models like Random Forest, Decision Tree, and Logistic Regression, identifies key predictors—including medication history, biological markers, and demographic factors—that influence psychiatric diagnoses. This work not only enhances our understanding of COVID-19’s mental health impact but also provides a foundation for targeted prevention and early intervention strategies, ultimately aiding healthcare providers in improving patient outcomes.

Change in Percentage of Common Features Over Different Time Frames

Comparison of Probability Distributions for Real vs. Synthetic Data

Key Steps and Techniques:

Data Collection and Preparation: Conducted data cleaning and cohort creation, using techniques like propensity score matching to ensure reliable comparisons across COVID-positive and control groups.
Exploratory Data Analysis: Analyzed demographic, clinical, and biochemical data to identify initial patterns and key variables correlated with psychiatric diagnoses.
Machine Learning Models: Built and evaluated various models to predict the likelihood of developing psychiatric conditions, including:
- Random Forest: Achieved high accuracy in identifying key risk factors, leveraging the model’s feature importance capabilities.
- Logistic Regression: Used to interpret how demographic and clinical features increase or decrease risk, providing a clear, probabilistic understanding.
- Decision Tree and Support Vector Machine (SVM): Utilized to compare model performances, visualize decision paths, and refine predictive accuracy.
Survival Analysis: Implemented Cox Proportional Hazards regression to measure the temporal risk of psychiatric outcomes, highlighting significant predictors across multiple time frames.

Top 10 Variable with the Greatest Difference of Features between Two SSPD Groups

Top 10 Feature Importances for Different ML Models

Confusion Matrix for Different ML Models

Top 20 Variables with Highest Percentage of 1

Presentation Link

Full Research Paper Link

Project Name: Predicting Suicide Risks by Leveraging Machine Learning Techniques.

Summary: This project aims to address the critical challenge of predicting suicide risks by leveraging machine learning techniques. By analyzing survey data, the study focuses on understanding key predictors of suicide risks and improving prediction accuracy. To tackle the issue of imbalanced datasets, data augmentation techniques like synthetic data generation were implemented, ensuring robust model performance and reliable insights.

Machine learning algorithms, including classification models and feature selection techniques, were utilized to identify significant predictors and enhance predictive accuracy. The goal of this project is to provide actionable insights that can guide targeted mental health interventions, enabling proactive measures to support at-risk individuals.

Open GitHub Repository ▶️

Project Name: BreastScanNN: Deep Learning for Breast Cancer Classification.

Summary: Built a neural network using TensorFlow/Keras to classify breast tumors as benign or malignant from 30 structured clinical features. The project demonstrates deep learning applied to tabular medical data rather than imaging — a less common but practical use case — with a clean Sequential architecture and a full training/evaluation pipeline.

View in Github 🛎️

Project Name: Heart Failure Mortality Prediction Using Machine Learning.

Summary: Compared six classification algorithms — including Random Forest, SVM, and Logistic Regression — on clinical records from 299 heart failure patients to predict mortality risk. The project emphasizes clean, replicable preprocessing and side-by-side model evaluation, demonstrating how data-driven risk stratification could support earlier clinical intervention.

View Project on GitHub📍

Project Name: Diabetes Risk Prediction in Female Patients Using SVM.

Summary: Built a support vector machine classifier to predict diabetes risk from clinical indicators (glucose, BMI, blood pressure) in a female patient cohort. The project walks through a complete, interpretable ML pipeline — from standardization and stratified splitting to performance evaluation — tailored to a specific demographic often underrepresented in general diabetes datasets.

View in Github 🔗

Project Name: Heart Disease Prediction Using Logistic Regression.

Summary: Developed a logistic regression model to predict heart disease presence using clinical and demographic features such as cholesterol, blood pressure, and chest pain type. The project focuses on identifying interpretable, high-risk predictors to support earlier diagnosis and intervention in a leading global cause of mortality.

View in Github 🎯

Project Name: Machine Learning Decision Support Tool for Differentiating Malignant vs. Benign Soft Tissue Masses on MRI.

Summary: Built a decision tree classifier (scikit-learn) on 295 retrospectively reviewed musculoskeletal MRI cases to help non-expert clinicians distinguish malignant from benign soft tissue masses, achieving 86% accuracy and an AUC of 0.84.

Identified lesion size (>4.2 cm), T1 heterogeneity, and perilesional edema as the strongest predictors of malignancy, with odds ratios up to 46.6x for perilesional edema (95% CI: 21.6–100.3, p<0.001).

Validated inter-reader reliability using Cohen's Kappa (substantial agreement, κ=0.70) and quantified diagnostic performance with sensitivity (75%), specificity (93%), PPV (85.9%), and NPV (86.6%) metrics.