The Critical Need for Trustworthy AI: Unpacking Fairness in Alzheimer's Disease Prognosis Models

      Alzheimer’s disease (AD) presents one of humanity's most profound healthcare challenges, impacting millions globally. This progressive neurodegenerative disorder, characterized by irreversible cognitive decline and loss of independence, affects a significant portion of older adults, particularly those aged 85 and above. A critical hurdle in Alzheimer’s care is the inherent unpredictability of its progression. Individuals, even with similar clinical profiles, can experience vastly different rates of decline, complicating clinical decision-making, long-term planning, and patient counseling. Traditional assessments often identify AD only after substantial neurological damage has occurred, limiting the effectiveness of early interventions and potential disease-modifying therapies.

      This challenge underscores an urgent need for advanced predictive frameworks. Rather than merely detecting AD after its onset, the medical community requires models that can forecast when clinically meaningful transitions are likely to occur. This necessity aligns perfectly with the principles of survival analysis, a statistical method designed to predict the time until a specific event. For a degenerative condition like AD, where disease processes unfold over extended periods, survival analysis offers a powerful lens to estimate an individual’s risk trajectory and support proactive intervention.

Deep Learning for Predicting Disease Progression: The Promise of Survival Models

      Recent breakthroughs in deep learning (DL) have revolutionized many high-dimensional tasks, extending their impact into medical prognostication. These advancements have enabled researchers to apply sophisticated AI models to predict disease progression in various critical domains, including cancer, embolism, relapse, and increasingly, Alzheimer's disease. Unlike traditional classifiers that offer a binary (yes/no) diagnosis after progression, deep learning-powered survival models can estimate an individual's unique risk profile, such as the likelihood of converting from a non-AD state to an AD diagnosis over a specific timeframe.

      Specifically, nonparametric deep survival models (NDSMs) have emerged as highly effective tools. These models can generate individualized survival distributions (ISDs) without imposing restrictive assumptions about the underlying data patterns. This is crucial for conditions like AD, where patient heterogeneity is significant. By breaking down the disease progression timeline into discrete intervals, NDSMs estimate the probability of an event (like disease progression) occurring at each interval, offering a far more nuanced understanding of an individual’s journey. While some traditional methods exist, the ability of NDSMs to model survival probabilities at a fine-grained, individual level often leads to superior performance in predicting time-to-event outcomes.

Beyond Accuracy: Why Trustworthiness Matters in Healthcare AI

      While the predictive power of NDSMs in survival modeling is impressive, merely focusing on performance metrics like the "Concordance Index" (which measures how well a model ranks individuals by their predicted time to event) falls short, especially in sensitive domains like healthcare. A truly robust and responsible AI framework in clinical settings must extend beyond simple accuracy to encompass the concept of "trustworthiness."

      Trustworthiness in AI for healthcare means ensuring models are not only performant but also fair, interpretable, and well-calibrated. Fairness ensures that predictions are unbiased across different demographic groups, preventing marginalized populations from receiving inequitable care. Interpretability allows clinicians to understand why a model made a specific prediction, fostering confidence and enabling informed decision-making. Calibration refers to how well the predicted probabilities align with actual observed outcomes. Alarmingly, many existing studies on deep learning in AD progression have primarily emphasized discrimination, often overlooking these critical facets. This gap makes it unclear whether these advanced models retain their desirable properties when applied to complex, real-world clinical data.

Novel Metrics for Quantifying Fairness in Deep Survival Models

      Recognizing this critical void, a rigorous study by researchers from West Virginia University and the University of Aberdeen set out to comprehensively investigate trustworthiness in AD progression modeling. A significant contribution of their work is the introduction of two novel fairness metrics specifically designed for nonparametric deep survival models, addressing limitations of traditional metrics that may not adequately capture bias in time-to-event predictions.

      These new metrics provide a crucial quantitative lens for assessing bias:

• Time-Dependent Concordance Impurity (CI-td): This metric evaluates how "pure" or consistent a model's risk ranking is across different sensitive groups (e.g., male vs. female, different ethnicities, education levels) over time. If the model consistently misranks individuals from one group compared to others, or if its ranking quality significantly varies, the CI-td would highlight this impurity.
• Kaplan-Meier Fairness (KM-Fair Calibration): This metric assesses how well the model's predicted survival curves for various sensitive groups align with the actual observed survival curves for those groups (known as Kaplan-Meier curves). It's a measure of calibration fairness, ensuring that the model's confidence in its predictions is equally reliable for all populations, not just the majority.

      By introducing CI-td and KM-Fair Calibration, the study provides a robust framework to quantify and analyze learned bias, making it possible to systematically evaluate fairness in NDSMs applied to real-world clinical cohorts.

Unveiling Bias: Key Findings and Their Implications

      The study conducted an extensive analysis of bias in state-of-the-art deep survival models across sensitive attributes such as sex, ethnicity, and education. The findings revealed a critical insight: while deep learning-powered survival models are indeed robust tools that can significantly aid clinicians in AD care decisions, they often exhibit "considerable bias" with respect to these protected characteristics.

      This discovery has profound implications for the deployment of AI in healthcare. Biased predictions could lead to unequal access to care, suboptimal treatment plans, or delayed interventions for certain demographic groups. For example, if a model systematically underestimates the risk of AD progression in one ethnic group due to historical data imbalances, individuals from that group might miss crucial early interventions. Understanding and mitigating such biases is not merely a technical challenge but an ethical imperative, essential for ensuring equitable and patient-centered healthcare. Technologies like ARSA AI Box Series and AI Video Analytics, while developed for different applications such as safety monitoring and traffic analysis, exemplify how AI must be rigorously evaluated for fairness and reliability before deployment in any sensitive environment.

Building Trustworthy AI: A Path Forward for Alzheimer's Care

      The findings from this research underscore the necessity of a comprehensive evaluation pipeline for AI trustworthiness in healthcare. This pipeline must systematically analyze multiple dimensions, including discrimination (predictive accuracy), calibration (reliability of probability estimates), fairness (absence of bias across groups), and interpretability (understanding model decisions). The study also highlighted the importance of thorough feature importance analysis to identify the patient characteristics that are most crucial for reliable AD predictions, paving the way for more targeted and equitable model development.

      Developing AI solutions that meet these stringent trustworthiness criteria is paramount for their successful and ethical integration into clinical practice. This research serves as a vital step towards developing more equitable and effective AI solutions, ultimately aiming to enhance human capability in managing complex diseases like Alzheimer's. ARSA Technology is experienced since 2018 in delivering production-ready AI and IoT systems across various industries, prioritizing precision, scalability, and measurable impact, while also advocating for human-centered innovation and ethical AI deployment.

      The insights from this study provide a strong foundation for future research, urging developers and practitioners to integrate fairness and trustworthiness as core considerations from the initial design phase through to the deployment and ongoing monitoring of AI systems in healthcare.

      To explore how intelligent solutions can enhance your operations while upholding the highest standards of ethics and reliability, contact ARSA for a free consultation.

      Source:

      Thrasher, J., Heintzelman, K., Martone, P., Kotlowski, D., Bhattarai, B., Adjeroh, D., & Gyawali, P. (2026). Investigating Trustworthiness of Nonparametric Deep Survival Models for Alzheimer’s Disease Progression Analysis. Proceedings of Conference on connected Health: Applications, Systems, and Engineering Technologies (IEEE/ACM CHASE ’26). https://arxiv.org/abs/2605.04063