An Explainable Digital Biomarker Framework for Early Dyslexia Identification

For decades, the path to identifying dyslexia in children has followed a frustrating pattern known as "wait-to-fail." A child struggles with reading, teachers and parents watch for months or years hoping the difficulty will resolve itself, and only after the child has fallen significantly behind and suffered repeated academic embarrassment is a formal diagnosis made. This delay happens not because early signs are invisible, but because the tools to spot them are often too complex, too expensive, or too opaque to be used effectively in schools and clinics. We know that specific cognitive abilities, such as how quickly a child can name familiar objects or how well they can isolate individual sounds in words, are strong indicators of reading risk long before formal instruction begins. The challenge has never been a lack of evidence, but rather the absence of a simple, trustworthy instrument that can translate those early signs into a clear, actionable answer for a family.

This paper addresses that gap by designing a new framework for an early dyslexia identification tool, one built on the principle that a medical or educational decision must be understandable to the people making it. The researchers, working as part of a larger five-phase research program, have moved past the question of which signs to look for—which was settled in a previous phase to include four specific cognitive markers—and focused entirely on how to combine them. The goal was to create a system that does not function as a "black box," where a computer spits out a risk score without explaining why, but instead acts as a transparent guide. The resulting design, called the Explainable Digital Biomarker for Dyslexia (EDBD), is a specification for a fifteen-to-twenty-minute digital assessment that a child can take largely on their own. It takes the results from four distinct tasks and weaves them into a single probability estimate, but crucially, it also provides a plain-language explanation of exactly which factors drove that estimate, ensuring that a parent and a specialist are looking at the same underlying reasoning.

The core of this new framework is a deliberate choice to avoid the complex, hidden algorithms often used in modern artificial intelligence. While powerful machine-learning models can sometimes achieve high accuracy, they often operate in ways that are impossible to trace, leaving users to trust a number without understanding its source. The author argues that for high-stakes decisions involving a child's education and health, this opacity is unacceptable. Instead, they selected a simpler, additive architecture. Imagine this system not as a mysterious machine, but as a scale where each of the four cognitive markers adds a specific amount of weight to the final result. The system calculates a risk score by adding up the contributions of phonemic awareness (the ability to manipulate speech sounds), rapid automatized naming (how fast a child can name familiar items), orthographic choice (recognizing correct spellings), and verbal working memory (holding a sequence of words in mind). Because the math is additive and transparent, the system can show exactly how much each of these four areas contributed to the final score, rather than hiding that logic inside a complex neural network.

To ensure the results are trustworthy, the framework includes built-in safeguards for uncertainty. The system does not just give a single number; it provides a range to show how confident it is in that estimate, acknowledging that no test is perfect. It also checks the quality of the session itself. If a child is distracted, if the audio recording is poor, or if the child does not finish the tasks, the system flags the result as low quality and may withhold a specific probability number entirely, offering instead a recommendation to try again. This prevents the dangerous situation where a parent receives a confident-sounding number based on messy or incomplete data. The design also accounts for the fact that spelling rules vary across languages; while the default settings are calibrated for English, the framework includes a mechanism to adjust the weight of spelling-related tasks for languages with different writing systems, ensuring the tool remains fair and accurate across different regions.

The output of this system is designed for two different audiences, both looking at the same data but seeing it in a way that suits their needs. For families and educators, the result is presented as a simple visual grid of icons, where shaded icons represent the likelihood of dyslexia, paired with a short, plain-language sentence explaining the main reason for the score. For clinicians, the same data is shown as a detailed chart breaking down the contribution of each of the four markers, allowing for a deeper professional review. This dual-view approach ensures that the explanation given to a parent is not a simplified lie or a vague guess, but the exact same reasoning the specialist sees, just presented in a more accessible format. The framework also includes clear guidance on what to do next based on the score: a low risk suggests continuing with standard screening, a moderate risk suggests intensified support within the classroom, and a high risk suggests a referral for a full formal evaluation.

It is important to note that this paper does not present a finished, tested product. The author is offering a complete, detailed blueprint for a system that has not yet been built or piloted with real children. Every number in the design, from the initial weights assigned to each marker to the specific thresholds for risk levels, is provisional and waiting to be tested against real-world data in the next phase of the research. The author explicitly states that they have not yet proven that this specific design works better than other methods; rather, they have constructed a rigorous, evidence-based specification that avoids the pitfalls of unexplainable AI. They have mapped out a path where every design choice is traceable to existing scientific literature, creating a system that is ready to be built, tested, and refined. The ultimate goal is to turn the conviction that no child should struggle with reading for years without understanding why into a concrete, testable reality, bridging the gap between early scientific evidence and the practical tools needed to help children succeed.