A machine learning approach to predict amyloid and tau positivity using clinical features

Daniel Arnold, Wyllians Vendramini Borelli, Luiza Santos Machado, Nesrine Rahmouni, Joseph Therriault, Stijn Servaes, Jenna Stevenson, Arthur C. Macedo, Artur Francisco Schumacher Schuh, Christian Mattjie, Rodrigo C. Barros, Marco Antônio De Bastiani, Firoza Z Lussier, Mira Chamoun, Gleb Bezgin, Andréa Lessa Benedet, Tharick A. Pascoal, Pedro Rosa‐Neto, Eduardo R. Zimmer, Alzheimer's Disease Neuroimaging Initiative (adni)Alzheimer 2024 · DOI

TL;DR

This paper presents a machine learning model that can predict amyloid and tau positivity in individuals using only readily available clinical features. This approach could significantly improve the efficiency and reduce the cost of clinical trials for Alzheimer's disease by better selecting participants.

What Problem Does This Paper Address?

Clinical trials for anti-amyloid drugs for Alzheimer's disease often face a significant challenge: a high rate of "screen failure." This occurs when potential participants are found to not have amyloid pathology, even if they have cognitive decline. This is problematic because these drugs are designed to target amyloid, and enrolling individuals without it is a waste of resources and time. The authors propose that by leveraging existing clinical information, they can better identify individuals who are likely to have amyloid and tau pathology, thereby reducing screen failures and streamlining the trial recruitment process.

Methods

The researchers employed a machine learning approach, specifically using an XGBoost model. They selected common clinical features that were shared across multiple datasets, including age, Montreal Cognitive Assessment (MOCA) scores, clinical diagnosis (e.g., cognitively normal, mild cognitive impairment, dementia), sex, education, BMI, and the presence of certain medical conditions like heart disease, stroke, hyperlipidemia, and diabetes. The model was trained on a large dataset, with a portion allocated for training, validation, and testing. Feature importance was assessed using SHAP (SHapley Additive exPlanations) values to understand which clinical variables were most influential in predicting amyloid and tau positivity.

Datasets

The study utilized data from three well-established Alzheimer's disease research databases:

  • ADNI (Alzheimer's Disease Neuroimaging Initiative)
  • TRIAD
  • PPMI (Parkinson's Progression Markers Initiative)

This combined dataset comprised 1694 participants. The authors did not introduce a new benchmark dataset but rather leveraged existing, widely recognized resources for their study.

Key Insights and Ideas

  • Clinical data is surprisingly predictive: The study demonstrates that even basic clinical information, often collected during routine patient visits, can be highly effective in predicting the presence of amyloid and tau pathology. This challenges the notion that advanced neuroimaging or fluid biomarkers are always necessary for initial screening.
  • Age, Diagnosis, and MOCA are King: The SHAP analysis revealed that age, the individual's clinical diagnosis (CN, MCI, or dementia), and their MOCA score were the most significant predictors. This makes intuitive sense, as these factors are known to correlate with cognitive decline and the progression of Alzheimer's pathology.
  • Reducing trial costs: By improving the selection of participants for anti-amyloid trials, this machine learning approach has the potential to significantly reduce the financial burden and time investment associated with these crucial research efforts.

Key Results

The XGBoost model achieved a strong performance in discriminating between individuals with and without amyloid and tau pathology (A-T- vs. A+T+). The receiver-operator characteristic (ROC) analysis yielded an Area Under the Curve (AUC) of 0.86. The model demonstrated a sensitivity of 0.82, a specificity of 0.90, and an overall accuracy of 0.88. The top three most impactful features identified by the SHAP analysis were age, diagnosis, and MOCA score.

Limitations and Future Work

The authors acknowledge that this approach is a step towards improving participant selection for anti-amyloid trials. They suggest that a "two-step workflow" incorporating this machine learning model could be implemented. Future work could involve developing an "individualized calculator" based on these clinical variables, allowing for more personalized risk assessment and potentially wider adoption in clinical practice. Further validation of the model on independent datasets and in diverse populations would also be beneficial.

Summary updated

Abstract

Abstract Background Screen failure due to amyloid negativity is yet a problem in clinical trials for anti‐amyloid drugs. In this context, clinical characteristics of patients presenting with cognitive decline may decrease the screen failure ratio by increasing the odds of selecting individuals with brain amyloid pathology. Herein, we aimed at estimating amyloid and tau positivity in individuals using clinical variables in a machine learning model of prediction. Method We selected 1694 participants with amyloid and tau status from ADNI, TRIAD, PPMI databases. Commonly shared clinical features selected between datasets were age, total MOCA scores, clinical diagnosis, sex, education, BMI, heart disease, stroke, hyperlipidemia and diabetes. Amyloid positivity was defined by amyloid‐PET (PIB‐PET, FBB‐PET or AZD4694‐PET) or CSF AB42 and Tau positivity defined by Tau‐PET (MK6240‐PET or AV1451‐PET) or CSF p‐tau181. The dataset was split into training (49%), validation (21%), and testing datasets (30%). A XGBoost model was tuned, and then used to predict the tau status outcome in the testing dataset. Result A total of 927 men and 767 women were included with mean MOCA 25.7 ± 3.8 scores (Table). The population consisted of 1276 CN, 290 MCI and 128 dementia individuals (1357 A‐T‐ and 337 A+T+). The receiver‐operator characteristic (ROC) analysis showed that the area under the curve (AUC) was 0.86 for discriminating A‐T‐ vs. A⁺T⁺, with 0.82 as sensitivity, specificity as 0.90 and accuracy as 0.88. The top 3 most impactful features were found to be age, diagnosis and MOCA score through a SHAP value analysis (Figure). Conclusion Predicting amyloid and tau positivity with clinically collectible variables may improve selection of individuals for anti‐amyloid trials. A two‐step workflow using this approach may significantly reduce the costs of AD trials. The following studies may incorporate an individualized calculator using clinical variables to estimate amyloid and tau positivity.
BibTeX
@inproceedings{arnold2024machine,
  author = {Daniel Arnold and Wyllians Vendramini Borelli and Luiza Santos Machado and Nesrine Rahmouni and Joseph Therriault and Stijn Servaes and Jenna Stevenson and Arthur C. Macedo and Artur Francisco Schumacher Schuh and Christian Mattjie and Rodrigo C. Barros and Marco Antônio De Bastiani and Firoza Z Lussier and Mira Chamoun and Gleb Bezgin and Andréa Lessa Benedet and Tharick A. Pascoal and Pedro Rosa‐Neto and Eduardo R. Zimmer and Alzheimer's Disease Neuroimaging Initiative (ADNI)},
  title = {{A machine learning approach to predict amyloid and tau positivity using clinical features}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.091844},
}
Figure