RecallX

Inspiration

Alzheimer’s disease affects over 55 million people worldwide and remains one of the most devastating and underdiagnosed neurodegenerative disorders. Traditional diagnostic methods like MMSE, CDR, and MRI are time-consuming, expensive, and require specialized expertise — making them inaccessible for many, especially in under-resourced areas. Inspired by the potential of AI and machine learning to revolutionize healthcare, our goal was to create an accessible, scalable, and interpretable diagnostic support tool for early Alzheimer’s detection using structured clinical data.

What it does

RecallX is an AI-powered screening platform that automates Alzheimer’s risk prediction by combining supervised machine learning models with an interactive chatbot. It conducts conversational MMSE and CDR assessments using the Perplexity Sonaar Pro API for reasoning and scoring open-ended responses. The collected data, alongside patient demographics, is processed by machine learning models trained on the OASIS longitudinal dataset to predict dementia risk, offering quick, reliable, and scalable early detection support.

How I built it

We began by preprocessing and standardizing the OASIS longitudinal dataset, analyzing clinical features like age, education, gender, MMSE scores, and CDR. Multiple machine learning models — Logistic Regression, Naive Bayes, KNN, and SVM — were trained and fine-tuned using GridSearchCV, with model performance evaluated via Accuracy, F1 Score, and ROC AUC. A conversational AI chatbot was built on Google Cloud Platform, integrated with the Perplexity API to handle MMSE and CDR assessments, reason through responses, assign cognitive scores, and relay results to the best-performing model (Logistic Regression) for dementia risk prediction.

Challenges I ran into

•Translating clinical protocols like MMSE and CDR into structured, conversational chatbot flows •Handling natural language variability while scoring responses consistently using Perplexity API •Ensuring balanced model performance in terms of sensitivity and specificity, crucial for clinical applications •Dealing with assumptions in different ML models, like feature independence in Naive Bayes and complexity in SVM •Balancing ethical concerns and data privacy while dealing with health-related conversational data

Accomplishments that I'm proud of

•Successfully automated MMSE and CDR testing via conversational AI with Perplexity API reasoning •Achieved strong model performance, with Logistic Regression leading in F1 Score and ROC AUC — critical for clinical reliability •Designed a scalable, interpretable machine learning pipeline using structured healthcare data •Demonstrated that ML applied to structured tabular data can rival more complex unstructured data-based approaches in diagnostic support •Built a functional prototype that paves the way for accessible, early dementia detection tools deployable in remote or under-resourced settings

What I learned

•How to effectively preprocess, analyze, and model structured healthcare datasets like OASIS for clinical classification problems •The trade-offs and assumptions underlying different ML algorithms and their impact on real-world diagnostic tasks •The importance of conversational AI reasoning and handling subjectivity in open-ended health assessments •The significance of balancing technological innovation with ethical and human-centered care in sensitive domains like healthcare AI •How comprehensive model evaluation (beyond accuracy) is essential when clinical decision-making is involved

What's next for RecallX

•Extend the chatbot’s language capabilities for multilingual and culturally sensitive screening •Integrate additional biomarkers, longitudinal data, and patient history for improved model precision

Built With

• client
• jupyter
• matplotlib
• notebook
• numpy
• oasis
• pandas
• scikit-learn
• seaborn