Clinical Lab Manager - News, Editorial and Products for the Clinical Laboratory
AI Is Broadening the Impact of Clinical Informatics

AI Is Broadening the Impact of Clinical Informatics

Advancing AI pathology and multi-omics data integration will expand the role of clinical informatics in health care

In an increasingly integrated technological world, laboratory diagnostics are following suit to take advantage of rapidly expanding computational capabilities. Where we once relied exclusively on the labors of pathologists for histopathological interpretation of patient tissues, we can now use enhanced computing power to overcome the limitations of manual approaches. The integration of large-scale automated and computational pathology techniques with expanding patient datasets will dramatically broaden the scope and application of clinical informatics, and thereby inform new diagnostic and treatment modalities.  

Pathologists traditionally use histopathology as a tool to confirm and characterize disease in patient tissue sections, making this technique a cornerstone of clinical diagnostics. Though effective in its own right, traditional histopathology is limited in efficiency and scalability due to the amount of labor required by the pathologist. Furthermore, this manual approach makes it more difficult to integrate histopathology with large patient datasets, and thereby limits its use for both research and clinical purposes. 

To overcome such constraints, artificial intelligence (AI) has been developed as a means to support pathologists and enable them to increase efficiency and accuracy, and to manage large multi-source datasets. 

AI in cancer diagnostics

In AI pathology, deep learning is used to train computer algorithms called convolutional neural networks (CNNs) to “think” like a pathologist, allowing programs to scan and characterize whole-slide images based on pre-designed parameters. Though AI in pathology remains a work in progress, its potential use is already apparent, particularly in cancer diagnostics. 

A team led by Dr. Cheng-Yu Chen at Taipei Medical University recently trained CNNs to more efficiently classify high-resolution images of biopsies from lung cancer patients. Chen’s group used AI to categorize approximately 10,000 non-small-cell lung cancers as adenocarcinomas or squamous cell carcinomas with high sensitivity and specificity. The program also highlighted suspicious areas for additional evaluation. Not only did this work showcase AI’s ability to accumulate large datasets, but it also highlighted how pathologists remain integral to the system, as they were needed to conduct initial program training and follow-up evaluations. There are many other examples of AI in cancer diagnostics; the continued push toward improved automated pathology signals a rising interest in incorporating AI into clinical diagnostics.

Expanding clinical informatics with big data

Despite AI’s rapid development, integration of big data into clinical informatics is still in its infancy. The advent of high-throughput tools and the availability of multi-omics datasets including genomics, transcriptomics, proteomics, and metabolomics has created an exciting runway for system-level analysis of different biological layers. The Cancer Genome Atlas (TCGA), a landmark cancer discovery program, was created in 2005 to curate a wealth of data associated with various cancer types. Initially established as a tool to explore underlying genomic changes in cancer, the program has since adapted to reflect clinical focus on data at the RNA and protein levels, and thereby facilitate more extensive classification of a patient’s phenotype. 

Though these efforts are most prevalent among cancer patients, multi-omics clinical informatics projects have expanded to other common conditions such as inflammatory bowel disease. The 1000IBD project, launched by the department of Gastroenterology and Hepatology of the University Medical Center Groningen in 2019, generated multi-omics datasets to accompany the unique pathological phenotypes of more than 1,200 patients. Moreover, the 1000IBD project incorporated additional patient data including environmental factors and drug responses, suggesting that multi-omics datasets can be integrated with existing patient metadata to provide a more comprehensive view of an individual’s health status. 

While bioinformatics provides a promising avenue to integrate multi-omics data with more classical diagnostics like histopathology, there are even more direct links already on the horizon. Computational integration of large molecular datasets typically fails to account for spatial heterogeneity of a tissue of interest. This is particularly evident in the tumor microenvironment, where changes in gene expression in a total RNA sample may overlook dramatic differences in expression profiles between invasive cancer cells and normal cells within a single biopsy. Novel spatial transcriptomics techniques provide gene expression data while preserving tissue architecture, thus allowing relatively seamless integration of histopathology and expression analyses. 

To that end, Satoi Nagasawa and colleagues at the University of Tokyo recently used spatial transcriptomics to analyze pathological sections of ductal carcinomas of the breast, and found that specific gene expression profiles predicted the likelihood of cellular progression to invasiveness. Thus, integrated pathology and –Omics approaches have the potential to create a molecular fingerprint for individual patients and their disease course, thereby informing better intervention strategies. As the technology continues to develop, these approaches will likely incorporate AI pathology to broaden the scope of such studies to larger patient cohorts with a variety of different diseases.

Obstacles and opportunities

Streamlining the clinical use of new patient data sources like AI pathology and multi-omics datasets through computational approaches is certain to present unique challenges. Many –Omics databases are still expanding and we are still learning how –Omics profiles correlate with patient health status. As our interpretation of different datasets improves, they will become more applicable to clinical informatics at the patient level. Moreover, a large network of professionals including clinicians, pathologists, administrators, and computational and data scientists will need to establish and maintain communication lines to facilitate collection and integration of patient data. Pathologists must collaborate with computer scientists to train and validate AI algorithms, and further communicate with administrators, clinicians, and data scientists to integrate additional datasets and ultimately develop comprehensive patient information. 

Despite inevitable growing pains throughout the expansion of clinical informatics, large-scale patient data integration will yield a wealth of opportunity. New collaborations between health systems and computational experts will dramatically change our understanding of patient health, both broadly and at the individual level. With the inclusion of new types of biological data that more extensively characterize individual patients, personalized medicine will become more practical and effective. Patients will have access to more individualized diagnostics and therapeutic strategies, and ultimately improved chances for better health outcomes