Editors: Liron Pantanowitz, MD, PhD, MHA, chair of the Department of Pathology and professor of pathology, University of Pittsburgh Medical Center, and Matthew G. Hanna, MD, vice chair of pathology informatics and associate professor, Department of Pathology, University of Pittsburgh Medical Center.
A multimodal and temporal foundation model for virtual patient representations at health care system scale
July 2026—Apollo is a health care-scale multimodal foundation model designed to integrate longitudinal clinical data into unified virtual patient representations. It was developed by the authors using data from a major U.S. health care system and trained on approximately 25 billion clinical records from 7.2 million patients that spanned more than three decades. The data set incorporated 12 medical specialties and 28 clinical modalities, including pathology images, laboratory testing, clinical documentation, medications, and structured electronic health record data. Apollo represents a shift from single-task computational pathology models toward longitudinal multimodal patient modeling. It contextualizes pathology within the entirety of a patient’s health care trajectory. The authors described the system as having a unified atlas of medical concepts, integrating more than 100,000 unique medical events into a shared representation space. They assessed these compressed patient embeddings across 322 clinical forecasting and retrieval tasks derived from an evaluation set of 1.4 million patients. Apollo was evaluated on 95 disease-onset prediction tasks, 78 disease-progression tasks, 59 treatment-response tasks, 17 treatment-related adverse event tasks, and 12 hospital operation endpoints. The model predicted disease onset up to five years in advance in certain settings. Among the reported examples, the system achieved an area under the receiver operating characteristic curve of 0.87 for predicting melanoma mortality and a balanced accuracy of 0.97 for predicting in-hospital dialysis dependence. The model also supported semantic retrieval tasks, allowing patient similarity searches using multimodal clinical inputs, including text and imaging queries. Furthermore, the study highlighted the increasing convergence of digital pathology, multimodal artificial intelligence, and longitudinal clinical analytics. Histopathology images were treated as one component of an integrated computational patient profile instead of as isolated diagnostic artifacts. This framework aligns with emerging efforts in computable medicine, wherein data across pathology and laboratory medicine, including genomics; radiology; and clinical narratives are jointly modeled to improve prediction, stratification, and treatment selection. The work also reflects a broader trend within computational pathology toward foundation models that can integrate image and non-image data streams at scale. Despite the scale and technical sophistication of the study, several considerations must be taken into account, including the fact that the work remains a preprint and has not yet undergone peer review. In addition, the model was developed using data from a large multi-site, albeit single, health care network, raising important questions about external validation, domain shift, demographic generalizability, and portability across institutions with differing laboratory workflows, scanner ecosystems, reporting conventions, and patient populations. Furthermore, as with many large multimodal AI systems, operational deployment would require careful governance surrounding explainability, regulatory oversight, bias mitigation, and data provenance. These issues are particularly relevant because recent pathology studies have demonstrated that AI systems may inadvertently rely on confounding clinicopathologic correlations rather than true biologic signals. Nevertheless, Apollo represents one of the largest attempts to date to create a unified multimodal patient representation model that explicitly incorporates pathology data with longitudinal clinical records. The study provides an important preview of how future AI systems may evolve from machine-based systems using narrow diagnostic algorithms into enterprise-scale multimodal clinical-intelligence platforms that integrate pathology into comprehensive patient-centered computational reasoning.
Zhang A, Ding T, Wagner SJ, et al. A multimodal and temporal foundation model for virtual patient representations at healthcare system scale. ArXiv. 2026. doi.org/10.48550/arXiv.2604.18570
Correspondence Dr. Faisal Mahmood at faisalmahmood@bwh.harvard.edu
Carbon footprint of going fully digital in surgical pathology
Digital pathology is inevitable and valuable, but it is not environmentally neutral. Fully digital pathology laboratories rely on energy-intensive data centers with continuous power and cooling demands, and they have a significant carbon footprint. Adding artificial intelligence to digital workflows will further amplify environmental costs. Given accelerating climate change and its direct health consequences, pathology laboratories going fully digital have an obligation to reduce emissions where feasible. The authors, based in France, conducted a study in which they demonstrated that sustainability must be treated as a design constraint, not an afterthought, when going digital in the surgical pathology laboratory. They looked at the carbon footprint of fully digitizing surgical pathology across France, a country with approximately 250 surgical pathology labs and a relatively low-carbon electricity mix dominated by nuclear energy. The study arose from a growing recognition that health care contributes five to 10 percent of global greenhouse gas emissions and that digital transformation may, paradoxically, increase emissions. Using national data from 2021, the authors modeled a hypothetical scenario in which all histological slides were digitized. Approximately 29 million glass slides per year would generate about 43 petabytes of whole slide images, requiring nearly 500 scanners and a large-scale digital infrastructure. Their analysis included scanners, image-management systems, desktop viewing tools, network transmission, and data storage at local and external data centers. Artificial intelligence was explicitly excluded from the study to isolate the baseline environmental impact of digitization alone. The study showed that a 100 percent digital pathology ecosystem would generate approximately 1,100 to 1,260 tons of carbon dioxide equivalent for three months of storage, rising to approximately 2,100 to 2,900 tons of CO2 equivalent for one year of storage, depending on the storage strategy. Data storage was by far the dominant contributor, accounting for up to 76 percent of emissions in long-term storage scenarios. Desktop tools, particularly high-resolution diagnostic monitors, were the second largest contributor, followed by scanners and image-management systems. At the level of a typical surgical pathology laboratory in France, full digitization would increase the annual carbon footprint by four to eight percent, an impact comparable to adding 13 to 26 ultralow-temperature (−80°C) freezers. This increase would occur on top of existing environmental burdens from consumables, such as reagents, cassettes, and slides, which were shown, in earlier studies, to dominate pathology’s carbon footprint. The authors identified several mitigation strategies that could have a significant positive impact: reducing whole slide image-storage duration, avoiding unnecessary data replication, decreasing file sizes, extending equipment lifespan, sharing scanners across institutions, and eco-designing information technology infrastructure. They concluded that while digital pathology is inevitable, noncritical full-scale digitization risks undermining health care climate goals unless sustainability is explicitly integrated into design and policy decisions.
Tilmant C, Trécourt A, Béchu C, et al; for the TEAP Group. What is the carbon footprint of a 100% digital pathology scenario in France? Pathology. 2026. doi.org/10.1016/j.pathol.2025.12.008
Correspondence: Dr. Cyprien Tilmant at tilmant.cyprien@ghicl.net