journal articles
2026 FIELD GUIDE TO SCIENTIFIC OPERATING SYSTEMS FOR LONGITUDINAL DISEASE RESEARCH: ARCHITECTURES, WORKFLOWS, VALIDATION FRAMEWORKS, AND EMERGING DIRECTIONS IN ML-ENABLED BIOMEDICAL RESEARCH
Ara S. Khachaturian, Nicolaas P. Pronk, Sudhir Sivakumaran, Jacques Touchon, Michael W. Lutz, Jakub Hort, Ornit Chiba-Falek, Ira R.J. Hebold Haraldsen, Stephen Z. Levine, Václav Snášel, Zaven S. Khachaturian, Jean-Marie Bouteiller
VM&E 2026;9:1-27
Decades of biomedical investment have produced vast repositories of imaging, molecular, genomic, and longitudinal clinical data. Yet the field’s capacity to translate this abundance into reliable predictions of individual disease trajectories remains deeply constrained. The principal obstacle is no longer one of data quantity; it is increasingly one of scientific architecture. Heterogeneous data streams remain fragmented across modalities and institutions, analytical workflows lack provenance and reproducibility, and most predictive models are evaluated in isolation rather than as components of a continuously learning scientific system.
The central scientific problem is therefore not simply the lack of data or algorithms. It is the absence of analytical frameworks capable of integrating and evaluating evidence across multiple biological scales, temporal horizons, and representational modalities while preserving scientific meaning.
This field guide examines how multimodal longitudinal disease modeling may require more than advances in AI models and algorithms. It may depend upon the deliberate construction of scientific operating systems: integrated systems spanning data, models, workflows, validation, governance, and regulatory alignment capable of ingesting heterogeneous data, preserving evidence provenance, supporting structured biological reasoning, and continuously refining predictions as new knowledge accrues. Rather than proposing a single implementation, this field guide surveys the emerging components of such a framework. The convergence of large multimodal foundation models, knowledge graph architectures, neuro-symbolic reasoning, and agentic AI systems now makes this aspiration technically plausible, but only if it is pursued as a coordinated systems-level effort rather than a collection of isolated modeling advances.
Six interdependent themes emerge repeatedly across the literature and current practice:
1. Data as scientific infrastructure. Dataset engineering, harmonization, annotation, and provenance tracking are not preparatory chores but first-class scientific activities.
2. Validation as a system-level property. Evaluation frameworks must incorporate temporal prediction, calibrated uncertainty, and real-world robustness rather than relying solely on conventional static performance metrics.
3. Modular and interoperable AI architectures. Foundation models, knowledge graphs, and agentic workflows should function as coordinated systems rather than isolated tools.
4. Workflow infrastructure as a scientific asset. AI-enabled pipelines should function as auditable, reproducible computational laboratory notebooks.
5. Precompetitive collaboration. Federated data access, shared benchmarks, and community-governed standards are essential for cumulative scientific progress across institutions.
6. Regulatory alignment by design. Regulatory frameworks, including FDA and EMA guidance, should shape traceability, lifecycle management, uncertainty quantification, and real-world validation from the outset.
The corresponding engineering challenge is the coordinated construction of the scientific infrastructure through which such frameworks become reliable, reproducible, and clinically actionable. Whether this opportunity is realized will depend not on any single breakthrough, but on the collective ability to build durable, interoperable, regulatory-aligned scientific systems through which predictive medicine becomes possible.
Scope and Intended Use
This document is intended as a periodically updated field guide for researchers, clinicians, engineers, and policy stakeholders working at the intersection of machine learning and longitudinal disease research. Its purpose is to survey emerging architectures, methods, infrastructure requirements, validation frameworks, and implementation challenges across a rapidly evolving field.
The document does not propose a single preferred architecture, technology stack, regulatory pathway, or commercial model. References to scientific operating systems, multimodal AI systems, foundation models, knowledge graphs, agentic workflows, and related concepts are intended to describe and organize the current landscape rather than establish standards of practice or make claims regarding superiority, readiness, or future adoption. Readers are encouraged to view this publication as a living resource that will evolve as technologies, evidence, and operational experience mature.
CITATION:
Ara S. Khachaturian ; Nicolaas P. Pronk ; Sudhir Sivakumaran ; Jacques Touchon ; Michael W. Lutz ; Jakub Hort ; Ornit Chiba-Falek ; Ira R.J. Hebold Haraldsen ; Stephen Z. Levine ; Václav Snášel ; Zaven S. Khachaturian ; Jean-Marie Bouteiller (2026): 2026 Field Guide to Scientific Operating Systems for Longitudinal Disease Research: Architectures, Workflows, Validation Frameworks, and Emerging Directions in ML-Enabled Biomedical Research. Vitality, Medicine & Engineering http://dx.doi.org/10.14283/VME.2026.1