🧠 Predicting ADHD with Machine Learning 

Data Science, Data Visualization, Preprocessing. Exploration, Machine Learning

🧠 Understanding ADHD Through Data Science

I'm excited to share the results of my final project for the CE888: Data Science and Decision Making module, where I tackled a real-world healthcare challenge: using machine learning to predict ADHD diagnosis and participant sex from neuroimaging and socio-demographic data.

💡 The Challenge

As a data scientist for the NHS (scenario-based), I built an explainable and fair predictive system to support clinicians in diagnosing ADHD — especially in female patients, who are often underdiagnosed.

🔍 Stage 1 – Data Exploration & Preprocessing

Cleaned and merged fMRI connectome data, socio-demographics, emotional, and parenting features for 1,213 participants.

Applied imputation techniques (e.g. IterativeImputer) and mutual information for feature selection.

Standardized numerical features and encoded categorical variables.

Used stratified train/validation/test splits to ensure balanced subgroup representation.

🧠 Stage 2 – Predictive Modeling & Fairness Evaluation

Trained several models: Logistic Regression, Random Forest, XGBoost, MLP, SVM, and KNN.

Evaluated using F1, Recall, and AUC-ROC with special attention to clinical relevance.

Random Forest was selected for ADHD prediction:

Test Accuracy: 82%

Recall (ADHD): 88%

AUC-ROC: 0.866

Applied LIME and SHAP for model explainability.

Conducted bias and fairness analysis across sex groups — ensuring high recall and balanced performance for both males and females.

📊 Tools & Libraries:

Python, Scikit-learn, XGBoost, Keras, LIME, SHAP, Pandas, Seaborn, Matplotlib

🎓 This project helped me sharpen my skills in:

-End-to-end data science pipeline development

-Interpretable ML in healthcare

Model fairness and explainability

Jupyter repository: https://github.com/DArmandoSalinas/Predicting-ADHD-sex/tree/main

🧠 Predicting ADHD with Machine Learning 

Data Science, Data Visualization, Preprocessing. Exploration, Machine Learning

🧠 Understanding ADHD Through Data Science

I'm excited to share the results of my final project for the CE888: Data Science and Decision Making module, where I tackled a real-world healthcare challenge: using machine learning to predict ADHD diagnosis and participant sex from neuroimaging and socio-demographic data.

💡 The Challenge

As a data scientist for the NHS (scenario-based), I built an explainable and fair predictive system to support clinicians in diagnosing ADHD — especially in female patients, who are often underdiagnosed.

🔍 Stage 1 – Data Exploration & Preprocessing

Cleaned and merged fMRI connectome data, socio-demographics, emotional, and parenting features for 1,213 participants.

Applied imputation techniques (e.g. IterativeImputer) and mutual information for feature selection.

Standardized numerical features and encoded categorical variables.

Used stratified train/validation/test splits to ensure balanced subgroup representation.

🧠 Stage 2 – Predictive Modeling & Fairness Evaluation

Trained several models: Logistic Regression, Random Forest, XGBoost, MLP, SVM, and KNN.

Evaluated using F1, Recall, and AUC-ROC with special attention to clinical relevance.

Random Forest was selected for ADHD prediction:

Test Accuracy: 82%

Recall (ADHD): 88%

AUC-ROC: 0.866

Applied LIME and SHAP for model explainability.

Conducted bias and fairness analysis across sex groups — ensuring high recall and balanced performance for both males and females.

📊 Tools & Libraries:

Python, Scikit-learn, XGBoost, Keras, LIME, SHAP, Pandas, Seaborn, Matplotlib

🎓 This project helped me sharpen my skills in:

-End-to-end data science pipeline development

-Interpretable ML in healthcare

Model fairness and explainability

Jupyter repository: https://github.com/DArmandoSalinas/Predicting-ADHD-sex/tree/main

ADHD Outcome Prediction – Exploratory Data Analysis & Preprocessing

Data Science, Data Visualization, Preprocessing. Exploration

🧠 Understanding ADHD Through Data Science

I recently worked on a project aimed at analyzing behavioral, demographic, and neurological data to explore patterns in ADHD diagnosis. This Stage 1 focused on data exploration and preprocessing, ensuring the dataset is structured for predictive modeling in Stage 2.

🔍 Key Insights from Data Exploration (EDA)

📊 Class Distribution & Imbalance

The dataset revealed a class imbalance, with more ADHD-diagnosed individuals than non-ADHD participants.

This means future models will need resampling techniques or alternative evaluation metrics to avoid bias toward the dominant class.

📈 Strong Correlation Between Behavioral Scores & ADHD

Features like SDQ Hyperactivity, SDQ Externalizing, and SDQ Difficulties Total showed a high correlation with ADHD outcomes.

These insights validate previous psychological studies linking hyperactivity and externalizing behaviors to ADHD diagnosis.

📉 Some demographic variables, such as MRI scan age, show low correlation with ADHD diagnosis, indicating they might not be strong predictors.

Further feature selection techniques will determine their significance in modeling.

🛠 Handling Missing Values & Outliers

The dataset contained minimal missing values, which were handled without significant data loss.

Outliers in behavioral scores were detected, but given that ADHD symptoms can vary significantly, these were retained to preserve data integrity rather than removed.

🚀 Preparing for Stage 2: Model Development

🔹 Feature Scaling

Models such as k-NN, SVM, and Neural Networks are sensitive to feature magnitudes, so Min-Max Scaling or Z-score Standardization will be applied in Stage 2.

🔹 Machine Learning Models Considered

Logistic Regression for baseline performance.

k-NN to leverage proximity-based classification on behavioral patterns.

Random Forest for handling mixed data types while capturing feature importance.

SVM to explore non-linear decision boundaries in high-dimensional data.

Neural Networks (if feasible) for detecting complex patterns in ADHD indicators.

🔹 Evaluation Metrics

AUC-ROC, Precision-Recall, and F1-score will be prioritized over accuracy to handle class imbalance effectively.

💡 Key Takeaways

✅ Data preprocessing is crucial—ensuring proper handling of missing values, categorical encoding, and feature scaling before model training.

✅ Feature correlation analysis provides valuable insights into which factors may contribute most to ADHD classification.

✅ Modeling strategies must account for class imbalance, as ADHD cases are overrepresented in the dataset.

Jupyter notebook in : https://github.com/DArmandoSalinas/ADHD-Outcome-Prediction-Data-Exploration-Preprocessing

Ready to move into Stage 2—model training and optimization! 🚀

Movie Recommendation System Using Machine Learning 

Artificial Intelligence, Machine Learning, Recommender Systems

Developed a comprehensive Movie Recommendation System using the popular MovieLens dataset to practice in machine learning, recommendation algorithms, and data preprocessing techniques. The system includes a simple web interface for user interaction, providing an intuitive platform for exploring recommendations.

Approaches Implemented:

1. Collaborative Filtering (SVD):

   • Built a User-Item matrix to identify latent patterns in user preferences through Singular Value Decomposition (SVD).

   • Predicted user ratings for unrated movies and recommended top movies for each user.

   • Achieved a Root Mean Squared Error (RMSE) of 0.477, significantly outperforming the baseline RMSE of 3.624.

2. Content-Based Filtering:

   • Leveraged TF-IDF vectorization and cosine similarity to recommend movies similar to a given movie based on genres.

   • Provided personalized recommendations for new users or sparsely rated datasets.

3. Hybrid Recommendation System:

   • Combined the strengths of collaborative and content-based filtering.

   • Adapted to varied user profiles by dynamically applying the most effective approach.

Web Integration:
Created a lightweight web application to showcase the recommendation system, built using Flask.

• Designed a user-friendly interface enabling users to input movie titles or user IDs for tailored recommendations.

• Demostrated integration of machine learning models with web technologies, ensuring efficient performance and scalability.

Competences

• Machine Learning and Advanced Algorithm Implementation (SVD, TF-IDF)

• Recommender Systems (Collaborative, Content-Based, and Hybrid)

• Web Development with Flask

• Data Preprocessing and Normalization

• Statistical Analysis and Performance Evaluation

Emotion Recognition Using XGBoost  

Artificial Intelligence, Machine Learning, Natural Language Processing (NLP)

Developed a text classification system for emotion recognition by analyzing tweets, utilizing the XGBoost algorithm for its exceptional performance in multi-class classification tasks. This project aimed to categorize tweets into six distinct emotional categories, showcasing the power of machine learning in processing social media data efficiently.

The project involved rigorous preprocessing techniques tailored to the unique challenges of tweet analysis. Steps included removing noisy elements like hashtags, mentions, and non-alphabetic characters, applying stemming to normalize words, and using Bag-of-Words for feature extraction. These processes optimized the dataset for modeling and ensured the system could handle over 400,000 entries effectively.

The XGBoost model delivered robust performance metrics, including an accuracy of 87.5%, by leveraging stratified K-fold cross-validation to ensure reliability and consistency. Visualizations, such as word clouds and bar plots, were used to provide deeper insights into the most common words and their associations with specific emotions.

Future improvements, including feature engineering with TF-IDF and n-grams to capture contextual relationships, as well as hyperparameter tuning, were identified to enhance the model's effectiveness further. This project highlights the critical role of advanced algorithms and NLP techniques in unlocking insights from unstructured social media data.

Competences

• Natural Language Processing (NLP)

• Machine Learning Model Development and Evaluation

• Text Data Preprocessing and Feature Engineering

• Advanced Algorithm Implementation (XGBoost)

• Cross-Validation Techniques for Model Robustness

• Analytical Thinking and Problem-Solving

Rossmann Sales Forecasting Using Feedforward Neural Networks

Artificial Intelligence, Deep Learning, Time Series Analysis 

In this project, I developed a predictive model for Rossmann drugstore sales using a Feedforward Neural Network (FNN). The goal was to forecast daily sales up to six weeks in advance to support effective staff scheduling and resource management.

Key highlights include:

• Data Preprocessing: Engineered features from temporal data (Year, Month, Week), competition metrics, and promotional activity durations. Missing data was handled effectively to ensure model robustness.

• Feature Engineering: Created new metrics like competition open duration and promo activity duration to enrich the dataset and improve predictions.

• Neural Network Design:
  Built a multi-layer FNN with 3 hidden layers (128, 64, 32 nodes) using ReLU activation and dropout for regularization.
  Optimized with Adam optimizer and Mean Squared Error (MSE) loss function.

• Optimization Techniques: Implemented early stopping to prevent overfitting and ensure generalization.
  Evaluation and Deployment: Preprocessed data was used to train the model and evaluate it on the Kaggle Rossmann dataset.

The project demonstrates the application of deep learning techniques to structured/tabular data, leveraging TensorFlow/Keras for implementation. It emphasizes efficient data preprocessing and feature engineering, which are critical for achieving reliable results in real-world machine learning tasks.

Competences

• Deep Learning

• Machine Learning

• Data preprocessing

• Feature Engineering

• Python Programming

• TensorFlow/keras

• Time Series Analysis

• Data Visualization

• Statistical Analysis

Autonomous Navigation System for TurtleBot 

Intelligent Systems and Robotics

Developed and implemented a fuzzy logic and PID-based autonomous navigation system for a TurtleBot. The system enables the robot to follow a predefined edge, avoid obstacles, and react to its environment in real time. Leveraged Python to design modular and structured code for sensor data interpretation, rule-based decision-making, and smooth robotic movement. The approach combined linear and angular velocity control to optimize performance.

Competences

• Analytical Thinking and Problem-Solving

• Technical Proficiency in Fuzzy Logic and PID Control

• Sensor Data Integration and Real-Time Processing

• Code Modularity and Organization

• Adaptability and Learning Agility

• Quality Assurance and Testing

CS50x Final Project

DIAR NOTICES

Software

I developed a digital platorm to recevive and send notices, naming it as a distribution canal of notices. For this project I used software developing skills. I programmed with python, JavaScript, HTML, CSS, Flask. 

Competences

• Analytical Thinking and Problem-Solving

• Technical Proficiency

• Project Management

• User Experience Design

• Adaptability and Learning Agility

• Quality Assurance and Testing