AI vs. Alzheimer's: How Machine Learning is Revolutionizes Clinical Trials for the Brain
The race to develop effective treatments for Alzheimer's disease (AD) is one of the most critical challenges facing modern medicine, but clinical trials for AD therapies face enormous hurdles. Alzheimer's disease progresses slowly, with neurodegeneration starting approximately 20 years before physical symptoms appear. This slow pace makes it difficult to measure whether an experimental treatment is actually working within the typical timeframe of a clinical trial, which is often less than two years. To demonstrate the impact of a disease-modifying treatment, researchers must be able to measure clinical decline within the study period. Unfortunately, data suggests that only about 40% of potential trial subjects will experience a measurable change in their clinical score, specifically a change in the Clinical Dementia Rating Sum of Boxes (CDR-SB) score within two years.
When clinical trials enroll patients whose condition is slow or non-progressing, it means the study needs a much larger sample size to detect the drug's effectiveness. This results in enrolling more patients who may not experience a measurable clinical benefit, leading to longer study durations and significantly higher costs. The traditional method of screening potential participants is itself a burden—it involves a series of procedures starting with medical history, moving on to cognitive tests, blood or spinal fluid analysis, and finally, expensive and specialized brain imaging. This whole process is time-consuming, invasive, and costly, and even after all that, it may still recruit patients who are unlikely to decline significantly during the trial period.
To overcome these obstacles and accelerate the development of new treatments for Alzheimer's disease, researchers have introduced a novel screening paradigm that incorporates powerful computing techniques, specifically machine learning (ML). This approach aims to dramatically improve trial efficiency by identifying the most suitable candidates for early-phase clinical studies—those patients who are most likely to show disease progression within the timeframe of the trial.
The Smart Funnel: Enhancing Screening with Machine Learning
The new screening paradigm is an innovative screening funnel that integrates ML-based models designed to predict disease progression. The core idea is to use ML algorithms to analyze massive amounts of patient data and identify patterns that pinpoint candidates most likely to progress quickly. This predictive power allows for a significant reduction in the time and cost associated with clinical trials.
This ML-enhanced funnel is structured as a staged approach that can be integrated into existing clinical trial workflows. The models are trained using comprehensive data, including neuroimaging, demographic details (like age), genetic markers (like APOE $\varepsilon 4$ status), and clinical assessments (like CDR-SB, MMSE, and ADAS13).
The screening process is divided into two major stages:
Stage One: Initial Assessment and Early Filtration
The first stage begins with a standard initial screening visit where fundamental information is collected, including demographics, clinical assessments, and genetic data. After passing traditional selection criteria, subjects are then filtered through the Stage One ML model. This model utilizes the initial non-imaging data (age, genetics, and clinical scales) to predict the risk of disease progression over 24 months. The aim of this stage is to identify and eliminate subjects unlikely to progress before they undergo expensive and burdensome imaging procedures.
Stage Two: Deep Imaging Analysis
Subjects who successfully pass the first stage are scheduled for a second screening visit to acquire specialized neuroimaging data, such as Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) scans. These images are processed using advanced ML techniques, specifically 3D convolutional neural networks (CNNs), to extract highly informative features related to disease progression.
The CNNs were designed to classify images based on current clinical status—distinguishing between healthy control subjects and those with dementia. When images from mildly impaired subjects are run through these pre-trained CNNs, the model outputs a probability score that serves as a proxy for disease progression risk; subjects whose brains resemble those of dementia patients are predicted to be faster progressors.
This neuroimaging risk score is then combined with the demographic, genetic, and clinical data from Stage One to train the Stage Two ML model. This final ensemble model determines which subjects should be enrolled in the clinical trial. By incorporating imaging data, the model's predictive power is significantly enhanced compared to traditional imaging features alone.
Results: Predicting Progression and Boosting Efficiency
The efficacy of the ML models was measured by their ability to predict subjects likely to experience faster clinical decline (defined as a 24-month CDR-SB increase of $\ge 1$).
The Stage One model, using only baseline demographic, genetic, and clinical assessments, achieved a good prediction accuracy with an Area Under the Curve (AUC) of 0.802. AUC is a measure of a model's ability to distinguish between progressors and non-progressors, where a score closer to 1.0 is better.
The Stage Two model, incorporating neuroimaging, showed significantly improved predictive performance. The best-performing single imaging modality model was the one incorporating amyloid PET scans, which achieved an AUC of 0.836. Models combining multiple imaging modalities (like amyloid PET, FDG PET, and MRI) did not show a significant improvement over the best single-modality model. Importantly, the results showed minimal overfitting, confirming the reliability of these performance metrics.
Quantifiable Savings in Trials
Applying this ML-enhanced screening funnel to simulated clinical trial recruitment workflows using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) revealed dramatic gains in efficiency compared to a traditional screening funnel.
The traditional screening funnel selected subjects who had an average 2-year CDR-SB change of 1.04. To detect a theoretical drug effect of 35% slowing of decline, this traditional funnel would require 424 subjects per arm.
In stark contrast, the ML-enhanced strategy selected subjects with a much higher average 2-year CDR-SB change of 1.71. Because these patients were more likely to progress, the power analysis showed that the study would only require 192 subjects per arm to detect the same drug effect.
This means that the ML approach could reduce the number of subjects required by 55%. Even though the overall screen failure rate (SFR) increased from 55% to 75% because the ML model is designed to be more selective, the substantial reduction in the total number of subjects required still resulted in fewer patients needing to be screened overall. The ML funnel required 180 fewer screening subjects per arm—a 19% reduction.
These efficiencies directly translate into time savings: the ML approach could shorten the amount of time required to complete study enrollment by approximately 13 months. Furthermore, the staged approach significantly reduced the need for expensive imaging procedures. The Screen Failure Rate for the amyloid imaging visit dropped substantially because the Stage One ML model filtered out unsuitable candidates early, resulting in a 72% reduction in screening amyloid PET scans.
Customizing Trials for Specific Goals
One of the key advantages of this enhanced screening funnel is its flexibility. Since the ML models output a probability score indicating progression risk, researchers can adjust the cutoff threshold to optimize the screening process based on the specific goals of their study. Three key optimization strategies were demonstrated:
Minimizing Screening Subjects (Fastest Recruitment): By setting thresholds to minimize the number of subjects who must undergo screening, researchers could further reduce recruitment time. This optimization resulted in a total time reduction of 19.5 months compared to the traditional funnel, making it ideal for studies struggling with recruitment or those requiring a short timeline.
Minimizing Study Subjects (Biggest Efficiency): Alternatively, thresholds can be set to recruit the absolute fastest progressors, minimizing the number of subjects required for enrollment. This strategy reduced the number of enrolled subjects by 55% compared to the original ML thresholds. Under this optimization, the number of screening amyloid scans could be reduced by 78% compared to the traditional funnel.
Selecting a Specific Progression Range: The model allows researchers to select subjects who fall within a specific range of expected clinical decline. This is crucial for studies where the treatment mechanism is hypothesized to be most effective only in patients progressing at a certain rate—for instance, excluding late-stage AD subjects with extremely rapid decline.
Context, Limitations, and Future Outlook
While highly promising, this ML-enhanced screening funnel is best suited for proof-of-concept or earlier phase trials where small sample sizes and rapid results are crucial, rather than large Phase 3 trials that require a wider, more representative patient population.
The researchers also noted current limitations. Since the model was trained primarily on historical ADNI data (a public-private partnership aimed at measuring progression of mild cognitive impairment and early AD), there are potential demographic biases that must be carefully considered during implementation. Additionally, the model needs to be validated on an independent dataset to confirm its performance. The current model did not include valuable biomarkers like tau PET or fluid biomarkers, such as phosphorylated tau, due to insufficient data availability, but incorporating these in the future is expected to further improve predictive power.
Despite these limitations, it is clear that machine learning offers significant benefits when integrated into patient recruitment for clinical trials. By reducing patient burden, shortening timelines, and dramatically lowering the number of required study subjects—potentially shortening study length by 13 to 19 months depending on optimization—this enhanced screening funnel could accelerate the development pipeline for AD therapies. Ultimately, the goal of making clinical trials faster and more efficient is to speed up the process of delivering lifesaving treatments to patients in need more quickly.
ML Scientists:
Carlos Guestrin
A leading computer science researcher, Guestrin is a Professor at Stanford University known for his work in scalable machine learning and algorithm development.
He is the co-founder of Turi, a machine learning platform later acquired by Apple, which helps developers create intelligent applications.
His research focuses on machine learning's potential to transform sectors like healthcare by enabling personalized, preventive care.
Laura Montoya
Montoya is the founder and executive director of Accel.AI, a global nonprofit that works on artificial intelligence initiatives for social good.
She launched the nonprofit LatinX in AI to create a network for Latino professionals in the field and to advance their research and projects.
Raised by a single mother in the San Francisco Bay Area, Montoya is dedicated to making AI education more accessible and ensuring that diverse perspectives are represented in the development of intelligent systems.
Lilian Rincon
A Senior Director of Product Management at Google, Rincon has been instrumental in the development of new features for the Google Assistant.
Born and raised in Venezuela, she previously worked at Microsoft, where she contributed to products like Skype and Bing.
As an expert in AI and machine learning, she was named one of the 15 most powerful women at Google in 2019 and is widely recognized as a major Latina leader in technology.
