RNN Neochord: Clinical Time Series Prediction

Status: ✅ Completed (UNIPD Master II) | Tech Stack: R, LSTM, Time Series Analysis

Problem Solved / Scientific Goal

The goal of this project was to create a predictive model capable of analyzing post-operative clinical time series. In medical contexts, data is not static: the progression of a pathology or response to treatment are dynamic processes over time that require models able to "remember" past dependencies.

Architecture Implemented

The project is based on a Machine Learning approach applied to healthcare data:

  1. Data Preprocessing: Cleaning and transformation of raw clinical data into structured time series.
  2. Feature Engineering: Identification of the most relevant clinical parameters for prediction.
  3. LSTM Modeling: Implementation of Recurrent Neural Networks (LSTM) to capture long-term dependencies in time series.
  4. Validation: Testing the model through cross-validation to guarantee robustness of predictive capability.

Results & Takeaways

  1. Temporal Management: I learned that the main challenge in clinical data is not just the quantity of data, but their sequence. LSTMs are a powerful tool, but they require extremely rigorous data preparation.
  2. Rigorous Validation: In the clinical context, model validation is more important than its architecture. Without proper cross-validation, the model risks overfitting on historical data not representative of reality.
  3. Clinical Integration: This project demonstrated how bioinformatics can provide concrete tools for decision support in pilot clinical pathways.

Tech Stack & Tools Used

  • R Programming Language
    • LSTM (Long Short-Term Memory) Time Series Analysis Cross-Validation

GitHub Repository

Check out the code: RNN_neochord on GitHub