Author's Draft: Published in the IEEE Computer Society flagship magazine Computer July 2025, pages 153-156.
Roy Rada
AI is transforming healthcare. Big-Data Radiation Oncology and High-Touch Palliative Care illustrate different opportunities.
Artificial Intelligence (AI) and healthcare have impacted one another since the beginning of AI. Traditionally, the branches of healthcare most dependent on digitized information benefited the most from AI. Do the latest advances in natural language processing impact the trajectory? Since Large Language Models (LLMs) exploit free-form text, do they open opportunities for branches of healthcare traditionally reliant on human-human interaction? Or will the areas that get the most investment already be the ones that get the most investment in the future, as a continuation of the famous aphorism 'the rich get richer, while the poor get poorer'.
The history of computer applications in healthcare supports the famous "Rich get Richer" phenomenon. In 1968 Merton introduced the 'Matthew effect' [1] as "This complex pattern of the misallocation of credit for scientific work must quite evidently be described as the Matthew effect ... For unto everyone that hath shall be given." Evidence for the Matthew effect has appeared across occupations. A study of healthcare delivery in China concludes, "the Intercity Health Network generates new structural inequalities in healthcare access exhibiting a Matthew effect [2]." New investments in digital information systems in healthcare have typically gone into areas that most relied on digital information [3]. Historically, the first healthcare computer investments were in accounting and next, radiology and pathology. Healthcare and AI illustrates the Matthew effect. This essay next reviews the AI and health literature before focusing on two application areas, radiation oncology and palliative care, which sit at opposite extremes of the capital spectrum.
Deep learning has been successfully applied to clinical medicine for more than three decades [4]. Generative AI trains on multimedia libraries, while Large Language Models (LLMs) train on document libraries [5], and both perform well on many health care tasks [6]. While LLMs lack epistemic validity, they create new opportunities for healthcare applications that converse with humans
A search of IEEE Xplore for ('artificial intelligence' AND ('health care' OR medicine) AND 'systematic reviews' AND 2020-to-2025) returned 54 citations. Most citations were for highly focused topics, such as "Multi-modal deep learning diagnosis of Parkinson s Disease" [8], but 9 were broad. One broad review shows that big data facilitates personalized medicine, predictive analytics, early illness diagnosis, precision diagnostics, and treatment optimization from electronic health records, genomes, wearables, and medical imaging [9]. In another broad review, adoption of AI applications depends on the eight factors of technical feasibility, ease of use, system quality, performance, usability, social influence, and trust [10]. The IEEE literature suggests that AI increases the effectiveness and efficiency of healthcare [11].
Large companies that develop electronic health records systems are working with large provider networks to integrate LLMs into the workflow.
Radiation oncology is one of three oncology branches, the other two being medical and surgical oncology. In preparing a radiotherapy treatment plan, the doctors acquire images, segment the target, and plan doses to kill cancer cells. Radiomics is a method to quantitatively analyze medical images to uncover tumoral patterns not appreciated by the naked eye and to support personalized therapy [12]. Doctors try to predict outcomes based on both radiomics and genomics. In the delivery of radiation, the team monitors patient changes and adapts treatment. Changing radiation based on patient motion occurs in real-time under computer control. Radiation oncology relies on massive amounts of digital information and repetitive tasks [13] -- 'Big Data' incarnate [14].
Radiation oncology is ripe for AI applications [15]: "AI ... transformative applications in radiation oncology given ... a heavy reliance on digital data processing and computer software." Radiation oncologists appreciate AI's impact and say that AI revolutionizes radiation therapy [16]. AI is transformative at every step of the radiation oncology process.
Palliative Medicine differs sharply from Radiation Oncology. The World Health Organization states [17]: "Palliative care ... improves the quality of life of patients and their families who are facing ... life-threatening illness, through the prevention and relief of suffering ... whether physical, psychosocial, or spiritual." Diseases benefiting from palliative medicine include cancer, heart failure, and stroke. Nevertheless, only 14% of people who need palliative care currently get it [17]. Health insurance underfunds palliative care, and hospitals, at a loss to themselves, provide more than half of the overall cost of palliative care [18]. A systematic review of challenges in palliative care concludes that the most significant barrier is a lack of resources [19]. If technology could reduce costs of delivering palliative care, then palliative care professionals have an obligation to explore using it [20].
Palliative care uses AI less than radiation oncology. Grant and patent activity evidences the divide. With the US National Institutes of Health (NIH) 'Research Portfolio Online Reporting Tools', users search a repository of both intramural and extramural NIH-funded research projects and access patents resulting from NIH funding [21]. A search on this NIH repository for 'radiation oncology' AND 'artificial intelligence' returns 68 active projects linked to 60 patents. A search on 'palliative care' AND 'artificial intelligence' retrieves 10 active projects and 0 patents. Radiation oncology is 'richer' than palliative medicine.
How are LLMs special for palliative care? LLMs can forecast mortality or monitor pain. For predicting mortality, one project uses traditional machine learning on structured data sets [22], while another project uses an LLM that reads unstructured medical records [23]. These two projects highlight the salient difference between LLM and non-LLM approaches. The LLM approach works on unstructured natural language. The LLM approach lends itself more readily to extensions and diffusion into Electronic Health Records systems.
An LLM for patient mental health showed that the LLM-system responds like a professional mental health counselor [24]. For the doctor, an LLM inferred wishes of mentally incapacitated patients based on prior records [25]. LLMs are well suited to palliative care [26].
One challenge in the roll-out of LLM applications is gaining the trust of the intended users. While this trust is a necessary condition for diffusion, another necessary condition is fitting into the workflow. The surge of interest in expert systems in the 1980s bares comparison to the surge of interest in LLMs. The 1980s MYCIN expert system was as good as expert physicians in its problem area of bacteremia, and it offered recency, accuracy, coherence, and transparency confirmed by clinical trials [27]. However, to use MYCIN, the physician needed to access a separate, stand-alone application and re-enter patient information into that application which took enough time that the cost/benefit ratio was too high for the physician -- in other words, MYCIN did not fit into the workflow [3]. Large companies that develop electronic health records systems are working with large provider networks to integrate LLMs into the workflow [28]. However, journals underreport these efforts.
Workflow systems rely on roles defined as functions with rules for passing messages among roles. LLMs can mediate between roles and between a role and the rules of that role [29]. In radiation oncology, automation of workflow with the support of LLMs is advanced [30], but not in palliative care.
A hospice care system, which is the ultimate example of palliative care, suits LLM integration into the workflow [31, 32]. The system would represent the roles of the interdisciplinary team and the roles of patients and their family, as 'dignity therapy' illustrates. Dignity therapy helps patients reflect on end-of-life issues. The patient receives nine standard questions which guide a conversation. The dialog is recorded, transcribed, and edited into a legacy document for the patient to use in implementing the last steps of life [33]. However, dignity therapy is underutilized due to a lack of trained staff to help. With an LLM-automated social worker, patients could develop their dignity document [34, 35].
Palliative care does not process vast amounts of information in real-time, as radiation oncology does. Palliative care relies on natural language communication, and LLMs facilitate natural language communication. As death approaches, the roles of palliative care and natural language communication increase. LLMs are generalists [36] and could support palliative care applications both clinician-facing and patient-facing across every phase of the clinician and patient experience.
The betting AI entrepreneur invests in radiation oncology rather than palliative care. Radiation oncology uses LLMs, and as LLMs are commoditized, their use diffuses into poorer domains, such as palliative care. The radiation oncology patient cannot hope to understand the intricacies of radiation treatment, trusts the doctor, and hopes for a cure. For the hospice patient, by contrast, the doctor is less relevant, and the patient could benefit from LLMs but gets no help in using LLMs. Would the betting person see anything different here?
Roy Rada is a Professor Emeritus at the Department of Information Systems, ITE Building, University of Maryland, Baltimore County, MD 21250. Contact rada@umbc.edu.