Approximately a decade ago, the signing of the HITECH Act by then President Barack Obama triggered a massive shift among healthcare providers to move from pen-and-paper to electronic health records (EHRs), and marked the advent of the digital healthcare revolution. The development of this digital foundation has had many hiccups along the way, but the benefits of having patients’ entire health records available at the click of a mouse have drastically improved the clinical process. Additionally, this revolution has created a plethora of data and medical images that now garner a greater need for analysis and interpretation.
The propulsive force driving artificial intelligence (AI) development is the realization that we have, in the developed world, an aging population facing growing incidences of chronic illness and morbidity, combined with unsustainable healthcare spending in dire need of curtailing. The question has become, “how can we do more, with a scarcity of resources, in an age of value-based care for patients?” Out of this dynamic, the impact of artificial intelligence in healthcare will be reduced.
AI is not a new concept; in fact, the term was first coined in 1956 by computer scientists and IBM employees at the Dartmouth Conference, and it refers to a machine or computer that exemplifies intelligence in executing a task. Deeper branches of artificial intelligence include machine learning and deep learning, in which the algorithm is learning based on pattern and data recognition. This is where deep neural networks are playing into the field of medical imaging. To create these image-analyzing algorithms, large amounts of medical images need to be gathered in a curated form and fed to the algorithm, so it can learn and decipher differentiating features. An example of this would be healthy physiological lungs, imaged via a chest CT, contrasted against lungs with nodule development.
The challenge in creating a reliable imaging algorithm is that the data being fed to it must be of a quality where variations in the images are mitigated. An example of this is an algorithm analyzing brain MRI scans from an image set with large variability derived from several factors — such as radiology technician variability, contrast agent variability, machine calibration variability, etc. In such a scenario, reliability of the image algorithm conclusions could be diminished. Another challenge for algorithm development is gaining access to images. The NIH has a publicly open repository, but many algorithm developing companies are formulating partnerships with healthcare facilities that can provide better quality images for the development of their solutions.
However, several voices in the industry have vocalized frustration with the trend towards viewing these images as an asset, thereby prompting some facilities to keep a tighter hold on this resource, which can consequently stifle and prolong the widespread penetration of commercialized solutions in the industry. Nonetheless, with the increased development and advancement of these neural networks, fewer images will be necessary to get the same accurate results.
Initial speculation in the industry was that radiologists would be the first clinician type to be superseded by the “rise of the machines.” This conjecture, much of which was perpetuated by healthcare media and personnel in the data science field who have a narrow conception of the radiology profession, has waned in favor of a more productive narrative stating that such technology will become a tool in the clinician’s armamentarium that can be employed to enhance the clinical workflow. The combination of machine and human input and interpretation strikes the right balance of intelligent augmentation for radiologist practice of the future. Areas, where this technology will have a profound effect, include 1) enhanced diagnosis and incidental findings of images, 2) predictive image analysis, and 3) workflow efficiency with data analysis, report generation, and prioritization of key findings are areas that offer a lucrative and unique selling point to a radiologist.
Superior AI algorithm solutions will differentiate themselves in a number of ways. They will create significant time savings with the generation of radiology reports, as well as highlight key findings. They will perform sophisticated tissue analysis and predictive analysis of medical imaging scans that cannot be done with the naked eye, and they will analyze all relevant information from the patient’s EHR history to provide a holistic individual profile for care.
Several players in this nascent market are pushing to create solutions and, consequently, varying challenges and opportunities have arisen. Large multi-national radiology PACS and diagnostic imaging vendors — such as Philips Healthcare, Siemens Healthineers, and GE Healthcare — are developing their own AI solutions to provide these products, though not exclusively, as supplementary or integrated offerings with the purchase of their capital equipment or HCIT products. Smaller companies that are not traditionally in the healthcare industry, or are spin-offs from academic research programs, make up another class of companies.
The general state of the market is that solutions being developed are narrow in scope and will remain this way for the foreseeable future. That is, many of the companies are developing algorithms that are focused on specific modalities, performing a specific scan, for a particular ailment, or some other distinct task. A notable example of this is lung nodule detection and quantification for CT and chest x-rays. Another exceptional and small group of companies is pursuing a holistic solution with the movement toward precision medicine, where they plan to employ algorithmic analysis not just for medical images, but also in genomic and pathological data. These companies’ goal is to amass large and diverse amounts of information and knowledge management for the most effective treatment program tailored to the individual.
Because most of the technology being developed is narrow in scope, the market is seeing a consolidation around some platform companies that are attempting to market themselves as variety “app stores,” from which algorithms can be chosen in an a la carte fashion or in bundled packages. Though these platform companies can be viewed in a class of their own, some of the large multi-national PACS and diagnostic imaging vendors developing AI are seeking to be platform solutions, as well.
This business model is a win-win for platform companies, algorithm developing companies, and the healthcare organizations that wish to begin employing some form of AI strategy in their imaging departments. Part of the reason is that many algorithm-developing companies are very much in the scientific stages of their endeavors, and thus do not have the luxury of large capital investment for scaling and expanding their business with a sales force to approach hospitals. This, compounded with the large up-front costs of initial installation and implementation of their AI products, lack of brand name recognition, and a lack of a historical track record in facilities, makes partnership with larger healthcare players a lucrative sales channel for the small AI-developing organizations. In addition, platform companies benefit because they have built their business models to be app stores that can approach hospitals with a wide range of algorithm solutions, which allows platform companies advantages with distribution. In the end, healthcare facilities benefit by not being confined to a restrictive selection of AI algorithms from a single vendor.
Furthermore, many strategic partnerships are being formed between different companies and vendors; examples include Agfa Healthcare and RADLogics, Zebra Medical Vision and Google Cloud. Increasingly, this market will see such partnerships as companies and vendors create strategies for harnessing the expertise of other players to increase the distribution, integration, and development of their AI solutions.
Discussions in the industry have made a point that the non-healthcare technology forces of Google, Amazon, Microsoft, Salesforce, and Apple may want to provide their own platform “app stores” for AI developers to sell through their channels. Other industry personnel believe these companies will continue to offer their API tools for developers to create their own algorithms, as well as offer their cloud services for image storage and analysis, staying out of the platform business model. Increasing FDA approvals in the next couple of years, accompanied with larger adoption rates of this technology in healthcare facilities, will formulate a more insightful picture of the direction the competitive landscape will take.
The need to increase the speed and value of the patient care process has been made increasingly evident by the healthcare burden in the United States. Per the Centers for Medicare and Medicaid Services(CMS), the United States currently spends approximately 18 percent of its national GDP on healthcare, an amount forecasted to grow to 20 percent by 2025. Artificial intelligence is one of many revolutionary technologies that will curb this spending behemoth. Today, as a society, we demand more of our healthcare; the driving goal is for us to live longer lives that defy the evolutionary paradox, enabling us to live these longer lives in a healthy state, avoiding chronic illness and morbidity.
AI will be an asset to the radiologist when it comes to workflow and diagnostic efficiency, as well as a vital tool that fosters a paradigm shift in predictive (as opposed to reactive) healthcare. Furthermore, it will enable the reallocation of healthcare resources where once resources consumed certain aspects of the system and created bottlenecks. Though this is a nascent market, and the technology has many hurdles to overcome related to reliability, user confidence, and adoption, movement towards AI is not something to be intimidated by, but something to embrace, as its enhanced potential is an essential ingredient for the future of value-based care.