Abstract
Neurodegenerative diseases such as Alzheimer’s and Parkinson’s impact millions of people worldwide. Early diagnosis has proven to greatly increase the chances of slowing down the diseases’ progression. Correct diagnosis often relies on the analysis of large amounts of patient data, and thus lends itself well to support from machine learning algorithms, which are able to learn from past diagnosis and see clearly through the complex interactions of a patient’s symptoms. Unfortunately, many contemporary machine learning techniques fail to reveal details about how they reach their conclusions, a property considered fundamental when providing a diagnosis. This is one reason why we introduce our Personalisable Clinical Decision Support System PECLIDES that provides a clear insight into the decision making process on top of the diagnosis. Our algorithm enriches the fundamental work of Masheyekhi and Gras in data integration, personal medicine, usability, visualisation and interactivity. Our decision support system is an operation of translational medicine. It is based on random forests, is personalisable and allows a clear insight into the decision making process. A well-structured rule set is created and every rule of the decision making process can be observed by the user (physician). Furthermore, the user has an impact on the creation of the final rule set and the algorithm allows the comparison of different diseases as well as regional differences in the same disease.
Introduction
Neurodegenerative diseases like Alzheimer’s (AD) and Parkinson’s (PD) are notoriously difficult to diagnose early, yet early intervention is critical for slowing progression. While machine learning (ML) excels at finding patterns in the massive, multi-modal datasets (imaging, genetics, clinical tests) associated with these diseases, most high-performing models are "black boxes." In a clinical setting, a "black box" is a liability. Physicians need to know why a system suggests a diagnosis to trust it and to explain it to patients. Furthermore, standard ML models are static; they don't easily incorporate a doctor’s "gut feeling" or specialized knowledge about regional variations in disease presentation. PECLIDES Neuro attempts to bridge this gap by providing an interpretable, interactive, and adjustable diagnostic tool.
The PECLIDES Pipeline
PECLIDES Neuro moves from raw data to a transparent rule set through a multi-stage pipeline designed for clinical utility.
- Random Forest Foundation: The system first trains a standard Random Forest classifier on patient data.
- Rule Extraction: It iterates through every tree in the forest, converting every branch into a logical "IF-THEN" statement (Disjunctive Normal Form). For example: IF (Age > 75) AND (ApoE4 = True) THEN Alzheimer’s.
- Rule Scoring & Personalization: This is the core innovation. The authors extend existing rule-scoring algorithms to include a personalization attribute. Users can nominate "favorite" features they consider highly relevant. Rules containing these features receive a score boost.
- Rule Set Minimization: To prevent "information overload," the system prunes the thousands of extracted rules down to a manageable few (often just 1-10% of the original set) based on accuracy, length, and user preferences.
- Interactive GUI: A graphical interface allows physicians to perform these reductions and view resulting rules in real-time.
Interpretability and Personalization
By converting complex forests into discrete logic, the system provides a "glass box" view. A physician can literally read the list of criteria the AI used to reach a conclusion. This transparency is vital for clinical validation.
Personalization isn't just about user preference; it's about clinical utility. If a doctor knows that a specific biomarker is particularly reliable in their specific patient population (e.g., due to ethnic or regional factors), they can "force" the AI to prioritize rules using that biomarker. The paper demonstrates that you can often discard 90-99% of the rules in a Random Forest and still maintain near-peak accuracy.
Experimental Results
The authors evaluated PECLIDES Neuro across several diverse datasets, including the Alzheimer’s Disease Neuroimaging Initiative (ADNI), PROPAG-AGEING, and specialized Parkinson’s datasets involving digitized spiral drawings and voice measurements.
- Clinical Data (ADNI): The initial Random Forest achieved 97% accuracy. After pruning the rule set by half, the system still maintained 94.7% accuracy.
- MRI Data: Using features extracted from 3D MRIs, the system achieved 99.38% accuracy in 10-fold cross-validation.
- Personalization Impact: On the MRI dataset, reducing the rule set to just 10% of its original size without preferences dropped accuracy to ~72%. However, when the user nominated three "preferred" features, the same 10% reduction maintained 100% accuracy. This highlights how expert guidance helps the algorithm retain the most clinically relevant rules during pruning.
Conclusion and Future Work
While the clinical data results are robust, the MRI evaluation used a relatively small sample (52 images), requiring validation on larger cohorts. Future work involves more sophisticated handling of missing data and the development of hybrid models like Deep Jointly Informed Neural Networks (DJINN). These models aim to combine the interpretability of decision trees with the predictive power of deep learning, potentially offering the best of both worlds for neurological diagnosis.
Summary updated