Исследователи выявили главное препятствие для клинического ИИ: данные, на которых обучаются алгоритмы, не отражают реальность болезни

Two peer-reviewed papers published within nine days of each other have independently reached the same conclusion: the most significant constraint on clinical artificial intelligence is not how the algorithms learn, but what they are given to learn from — and the data they receive was never built for the task. The more recent of the two, a viewpoint published September 1, 2026 in Frontiers in Science by Benjamin S. Glicksberg of the Icahn School of Medicine at Mount Sinai, argues that medical AI across every application domain remains bottlenecked by training data that is incomplete, compressed, and insufficiently representative of how disease and recovery actually unfold. The more technically specific of the two, a perspective paper published August 25 in Frontiers in Digital Health by South Korean researchers Prof. Eun Joo Yang of Seoul National University Bundang Hospital and Prof. Seung Hyun Chung of the National Cancer Center, goes further — naming the specific architectural reason this happens and proposing a framework for fixing it before any AI model is trained. Their paper, on clinical AI's representational bottleneck, introduces the concept of human-cognitive proxies and a design philosophy called Clinical Reality Translation-Construction (CRTC). The timing is pointed. Earlier this year, TechTimes reported on a landmark partnership between Korean telecom giant KT and Seoul National University Hospital specifically designed to solve hospital AI's data fragmentation problem — connecting incompatible electronic health record systems, radiology archives, lab databases, and nursing platforms into a shared AI-ready substrate. That infrastructure problem is real and documented. But the Yang-Chung paper argues there is a prior, deeper problem that no integration effort touches: even perfectly unified clinical data was built for human cognition and institutional administration, not for machine learning. Fixing interoperability still leaves AI training on data that systematically strips out the causal and temporal information that makes rehabilitation meaningful. How Clinical Scores Became AI Training Data (and Why That's a Problem) The records filling modern electronic health systems — diagnosis codes, functional assessment scores, clinical notes, performance ratings — are human tools built for human purposes. They exist to help clinicians communicate quickly, document encounters efficiently, bill correctly, and coordinate care across institutions. Yang and Chung call these records human-cognitive proxy clinical measures: compressed, institutionally legible snapshots of clinical reality that, by design, sacrifice mechanistic and temporal fidelity for speed, legibility, and administrative tractability. In rehabilitation medicine, this compression is especially visible. A Timed Up and Go (TUG) score records how many seconds it took a patient to rise from a chair, walk about 10 feet (3 meters), and sit back down. It does not record whether the patient used a hand grip to initiate the stand, how their trunk oscillated during the turn, whether their movement quality degraded over three successive attempts as fatigue accumulated, or whether they paused mid-task because of fear. A Fugl-Meyer Assessment for Upper Extremity (FMA-UE) score captures a point total across a structured sequence of arm and hand movements in a clinic setting — but two patients can reach the same score through entirely different mechanisms. One may have achieved genuine distal motor recovery in the affected limb. The other may have learned to recruit the opposite shoulder and trunk to complete the task through compensatory strategy. In proxy-based AI training, both patients look identical to the algorithm. Research using inertial sensors during the Box and Block Test — a standard upper-limb coordination assessment — found that over 80 percent of stroke survivors use substantial wrist or trunk compensation that conventional scoring cannot detect, with wrist compensation rates reaching 83.9 percent. The clinical score says these patients and their genuinely recovered counterparts are similar. The underlying movement science says they are not. What Medical AI Is Actually Doing: Second-Order Translation Yang and Chung describe current medical AI as "second-order translation." In the first stage, a clinician translates a complex clinical encounter into a human-oriented proxy: an image, a note, a code, a score, a label. In the second stage, AI interprets that proxy and generates another human-readable output: a risk score, a flag, a generated note, a recommendation. The AI has never interacted with the underlying clinical phenomenon at all — only with a compressed representation of a human's interpretation of it. This applies across the full range of clinical AI tools currently deployed. Diagnostic imaging AI learns from images that are themselves already shaped by acquisition practices, labeling conventions, and demographic biases. Clinical decision support systems draw on diagnosis codes and structured fields that reflect reimbursement logic as much as clinical reality. Natural language models trained on clinical notes learn the documentation habits of clinicians, not the underlying biology of their patients. The same structural problem applies to ambient AI scribes — the technology that has received significant industry attention for its ability to reduce physician documentation burden by listening to and transcribing clinical encounters in real time. However sophisticated, these systems generate output that slots back into the same proxy infrastructure the Yang-Chung paper critiques. More efficient documentation of a proxy-structured clinical system is still proxy-structured clinical AI. Glicksberg reaches a parallel conclusion from a different direction: much of the data health systems generate is discarded before AI can learn from it. The raw waveform behind an electrocardiogram interpretation. The synchronized video and electroencephalogram recorded during epilepsy monitoring. The high-frequency telemetry and kinematics of robotic surgical systems. Each is generated in real time, used for the task at hand, and then typically not retained — even though it encodes far more than the summary reaching the medical record. Why Cancer Patients Show the Gap Most Clearly The Yang-Chung paper uses cancer survivorship rehabilitation as its primary illustration, drawing on Prof. Yang's own clinical research program — and the clinical stakes are high enough to make the representational problem concrete. Chemotherapy-induced peripheral neuropathy — numbness and sensory dulling in the extremities — is a common and persistent side effect of several cancer drug classes including taxanes, platinum agents, and vinca alkaloids. Studies have found that around 47 percent of cancer survivors report ongoing neuropathy symptoms years after treatment completion, and that affected women face nearly twice the risk of a fall compared to those without neuropathy, according to CIPN and fall risk research synthesized from multiple cancer survivorship cohorts. The clinical pathway from chemotherapy to fall looks like this: peripheral neuropathy quietly degrades the body's ability to detect postural instability in real time. The patient compensates unconsciously but becomes aware of uncertainty on their feet. Fear of falling increases. Activity levels drop. Muscle deconditioning follows. Mobility declines. By the time this cascade reaches a clinical record, it may appear as a fall risk flag, a reduced six-minute walk distance, or a depression screening flag. The data captures the end states of the cascade. The pathway that connects chemotherapy to neuropathy to fear to deconditioning to functional decline — the chain that rehabilitation medicine actually needs to interrupt — is invisible. The clinical significance is not merely theoretical. A multicenter randomized controlled trial of a sensorimotor, movement-quality-focused intervention improved functional mobility and postural stability in breast cancer survivors without producing corresponding changes in grip strength, six-minute walk distance, or self-reported disability, as documented in Yang's multicenter breast cancer trial. The intervention worked. The conventional proxy measures used to evaluate rehabilitation outcomes could not see it working. An AI trained to predict which patients would benefit from such an intervention — using conventional outcome measures as its data source — would have had no basis for selecting the right patients. What the Framework Proposes Instead Yang and Chung call their proposal Clinical Reality Translation-Construction, abbreviated CRTC. It is not an algorithm, a sensor platform, or a new training technique. It is a design philosophy for how clinical data should be constructed before any AI model is trained — an upstream representational framework that asks a different question than existing clinical AI development typically asks, as laid out in their Frontiers Digital Health paper. Where current AI development asks "how should this model learn from available clinical data?", CRTC asks "what clinical entity should become available for learning in the first place, and what pathway through the patient's biology and behavior does it need to preserve?" The framework outlines a nine-stage design process: beginning with capture of the clinical situation through whatever data sources are available — wearables, video, audio, patient-reported outcomes, sensor arrays, clinical observation — and moving through selection of the relevant clinical phenomenon, segmentation of continuous events into meaningful units, modeling of the pathway through which a clinical problem develops, and representation in AI-readable form before any inference begins. The paper proposes six criteria for evaluating whether a representation qualifies as pathway-based rather than just richer-than-usual: mechanistic relevance (it connects to a plausible pathway generating the clinical problem); temporal trajectory integrity (it preserves not just that change occurred but the structure of how the trajectory evolved); contextual responsiveness (it changes meaningfully across task, fatigue, environment, and symptom context); clinical interpretability; additive explanatory or predictive value over existing proxies; and scale-explicit representation (it specifies what is captured and what is not, at which temporal and spatial resolution). CRTC does not replace existing frameworks for building clinical AI — digital phenotyping, digital biomarkers, computable phenotypes, digital twins — but operates at a layer upstream of all of them. Those frameworks typically start from a data stream, a validated feature, or a category definition. CRTC starts from a prior question: what clinical pathway should the data infrastructure be designed to make learnable? What Is a Pathway-Based Clinical Representation? To illustrate concretely: for post-stroke upper-limb dysfunction, the conventional proxy is a Fugl-Meyer score. A CRTC-informed representation would segment each reaching attempt into initiation, transport, grasp, and release phases, extracting trunk displacement, shoulder elevation, wrist stabilization, inter-joint timing, and fatigue-related degradation per trial. These features map to a "compensation-dominant vs. restoration-dominant" reaching phenotype — an AI inference target that predicts real-world arm use and therapy response in ways that a summary score cannot differentiate. The question shifts from "what is this patient's FMA score?" to "is this patient recovering genuine distal motor function, or consolidating a compensatory strategy that may limit long-term recovery?" Why Even Foundation Models and Ambient Scribes Don't Solve It A common response to concerns about clinical data quality is that modern foundation models and self-supervised learning techniques can extract richer representations from large datasets without relying on explicit human-designed labels. If you train a large enough model on enough EHR text, imaging data, and clinical notes, perhaps the latent structure of clinical reality will emerge. Yang and Chung address this directly. Modern representation-learning approaches, including self-supervised and foundation-model approaches, may derive latent representations without explicit proxy supervision — potentially capturing structure that labeled approaches miss. However, these methods can only access phenomena that data collection infrastructures have recorded. If rehabilitation data infrastructures consist primarily of scores, codes, and notes, even foundation models face representational constraints those infrastructures define. Glicksberg makes the same point about data density: text-only training is intrinsically low-density. Further scaling of text yields diminishing returns. Progress requires shifting toward higher-bandwidth data — continuous physiological monitoring, surgical video, longitudinal imaging — sources that carry the pre-symptomatic, longitudinal signal that clinically meaningful prediction requires. Data generated passively as a byproduct of documentation and billing, as Glicksberg argues on data density, "can capture neither the resolution nor the continuity that prediction requires." This has direct implications for clinical AI infrastructure investments. Building a shared data platform that unifies EHR systems and radiology archives — the goal of projects like the KT-SNUH partnership — is necessary for AI to reach clinical data at all. But it is not sufficient to make that data the right kind of data for AI to learn from. Integration delivers proxy-structured data more efficiently. The representational bottleneck, as Yang and Chung define it, operates one layer below integration. What Would Count as Evidence This Works The Yang-Chung paper is explicit that CRTC is a perspective framework, not a set of empirical findings. Its contribution is to define the representational problem, specify operational criteria, and outline testable hypotheses. The authors propose three concrete validation pathways. First, a transition-instability phenotype derived from wearable or video-based features during gait initiation, turning, and task transitions. Validation would test whether this phenotype predicts falls, mobility decline, or activity restriction better than conventional TUG or Berg Balance Scale scores. Second, a compensation-dominant reaching phenotype constructed from upper-limb kinematics — trunk displacement, shoulder elevation, inter-joint coordination, and fatigue-related degradation. Validation would test whether this phenotype distinguishes motor restoration from compensatory strategy and predicts real-world arm use more accurately than FMA-UE scores. Third, a symptom-function coupling phenotype derived from multimodal symptom reports, sensorimotor performance measures, and functional outcomes in cancer survivorship — tested against whether pathway-based representations better capture clinically meaningful change than proxy measures alone. The paper also acknowledges the costs and risks of richer representations. Kinematic profiles, ambient audio, continuous activity data, and symptom trajectories are privacy-sensitive in ways that a coded diagnosis is not. Richer data requires more complex infrastructure, device calibration, quality control procedures, and governance frameworks. Pathway construction itself may be influenced by investigator assumptions — different teams may construct different representations of the same clinical phenomenon. And increasing representational complexity does not automatically mean better representations; some pathway-based entities may prove difficult to operationalize within routine clinical workflows despite their theoretical value. The central claim of CRTC is testable: pathway-based clinical representations should provide explanatory, predictive, or actionable value beyond conventional proxy measures. If that value does not appear across empirical evaluations, the framework's practical justification would be substantially weakened. The authors do not claim otherwise. Though the paper is grounded in rehabilitation medicine — a domain where the representational gap between clinical scores and clinical reality is particularly visible — the argument extends. The same challenge arises in chronic pain, psychiatry, frailty, long COVID, and any other condition where dynamic adaptation, contextual dependency, and lived functional consequences are poorly captured by conventional data infrastructures, as Yang and Chung conclude. The paper was published open-access in Frontiers in Digital Health (DOI: 10.3389/fdgth.2026.1917999). Neither author received financial support for the work, and no conflicts of interest were declared. Frequently Asked Questions What is the "representational bottleneck" in clinical AI? The representational bottleneck is the gap between what clinical data was designed to do — help humans communicate, document, and bill — and what AI systems need it to do: encode the causal and temporal pathways through which clinical problems develop. Clinical records like Timed Up and Go scores, ICD diagnosis codes, and clinical notes efficiently compress a patient encounter into a human-legible form, but that compression strips out mechanistic information. An AI trained on these compressed records learns the administrative and documentation patterns of clinical practice, not the underlying biology. The bottleneck is "representational" because it exists at the level of how clinical phenomena are encoded, before any algorithm is applied — and fixing data integration or deploying better algorithms does not resolve it. Why doesn't training a larger AI model on more hospital data solve this problem? Even very large foundation models trained on vast quantities of clinical text, imaging, and records can only learn from the data they are given. If that data consists primarily of human-cognitive proxies — scores, codes, and notes built for human communication rather than for computational causal access — then a larger model will learn a richer version of the same proxy structure. Modern self-supervised learning approaches can extract latent representations without relying on explicit human-designed labels, but they still cannot recover information that the data infrastructure never collected. A model trained on summary TUG scores will not discover compensatory movement strategies that were never measured. Richer models improve only when the data itself is richer in clinically meaningful signal. What would pathway-based clinical AI data actually look like in practice? Rather than recording a summary score for how long a stroke patient took to complete a standing-and-walking test, a pathway-based approach would capture sensor data throughout the task: when exactly the patient initiated movement, how their trunk oscillated during the turn, whether their balance correction latency increased over successive trials as fatigue accumulated, whether they showed sensory substitution behavior. These measurements map onto mechanistically meaningful clinical entities — a "transition-instability phenotype" or a "compensation-dominant reaching phenotype" — rather than a single ordinal number. Such an entity would allow an AI to ask: is this patient building genuine motor recovery, or developing a compensatory strategy that works in the clinic but may not generalize to daily life? That distinction is clinically important and invisible to proxy-based AI. Does this critique apply to AI tools beyond rehabilitation medicine? According to the researchers, yes. The paper argues that similar challenges arise in chronic pain, psychiatry, frailty, long COVID, and any condition where dynamic adaptation, contextual dependency, and lived functional consequences are poorly captured by conventional clinical data. Chronic pain scales compress multidimensional experiences into numerical ratings. Psychiatric assessments compress behavioral states into diagnostic categories. These are clinically useful compressions, but they are not designed to give AI causal access to the processes they summarize. The researchers expect the representational bottleneck critique to apply wherever clinical AI is trained on data built for institutional communication rather than computational causal inference. ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.