Welcome to our first-ever whitepaper on the Triomics blog. Today, I hope to provide an overview of three key challenges facing the oncology community (sections 1-3 below) and, in section 4, their impact. In our fifth and final section, I discuss the role of generative AI companies like Triomics in helping solve these problems.
Any opinions are my own, and are shaped by my experience both as a physician and in digital health, including here at Triomics.
1. The explosion of cancer data
As the saying goes, quantity is a quality in of itself. The sheer amount of data in oncology is both staggering and ever-growing when examined through a multitude of lenses:
There are more cancer patients
The American Cancer Society (ACS) estimates that in 2024 there will be over two million new cancer cases. (All figures are U.S. only, unless otherwise noted.) In 2007, when the ACS first implemented their current methodology, they estimated there would be 1.45 million new cancer cases in 2008. The result is an estimated 36% increase in cancer cases over the last 16 years, or a 2.1% increase per year.
Patients live longer
The same ACS dataset estimates just over 610,000 cancer deaths in 2024, compared to just under 560,000 in 2007, meaning the death rate increased by 9.3% over the last 16 years, or 0.56% annually. While each cancer death is obviously tragic, that the death rate is climbing less than a third as quickly as the incidence rate means that the average patient is living longer with cancer. In turn, that patient is likely able to receive more treatments, see more doctors, have more labs drawn and participate in more trials–each of which generate more clinical notes and quantitative data.
Patients are sicker
To generalize the above point, patients live longer not only with cancer, but also with many other conditions. A JAMA Internal Medicine study this year quantified the resulting increase in medical complexity, finding that a hospitalized patient today had 2.7 times the odds of arriving via the emergency department than a patient 15 years ago, 1.5 times the odds of multimorbidity (having multiple significant medical conditions), 1.8 times the odds of polypharmacy (taking numerous drugs), and over double the odds of receiving treatment for multiple conditions during their hospital stay. (Despite all of this, today’s patients were actually slightly less likely to die during their hospital stay–a testament to medical advances.) In turn, this medical complexity generates more visits, labs and notes from other specialists that oncologists must consider. While this particular study analyzed patients in Vancouver, B.C., the findings certainly hold in the U.S., which grows slightly older and heavier each year.
There are more treatment options
Assume an oncologist had just five FDA-approved drugs to prescribe at the start of her career, but three new cancer drugs came to market each year thereafter. After a 30-plus year career, the oncologist retired some decades ago, when 100 cancer drugs had been approved. Intuitively, one might think the corpus of data related to cancer drugs increased 20-fold over her career, but this proportional increase in complexity actually understates the reality:
-First, the rate of drug discovery is increasing. In my example above, there was a steady rate of about three new drug approvals per year. In actuality, though, a 2021 JAMA study found a seven-fold increase in the rate of new FDA cancer drug approvals in the preceding decade-plus, as shown below. Presumably and hopefully this trend continues, as we consider the rate of discovery in immunobiology and genome editing (CRISPR), or the potential of AI to accelerate drug discovery.
Number of FDA-approved cancer drugs by year. Figure from Olivier et al, 2021.
-Second, complexity increases exponentially, even if knowledge were growing linearly. In my hypothetical, the number of possibilities would increase linearly if each patient were on just one of five drugs at the beginning, and just one of 100 drugs at the end of the oncologist’s career. However, how many treatment regimens would be possible (or how many drug-drug interactions must we consider) if each patient were on two such drugs? There are 10 choices (five choose two, if you recall high school combinatorics) for the initial patient, but 4,950 for the final patient. The nearly 500-fold increase in complexity from just a 20-fold increase in the number of drugs still understates matters, as we track more than two data points for each cancer patient, of course.
-Finally, many of these drugs work in new ways. About 50-100 of the drugs over these 11 years work in a new manner (a new “mechanism of action”), which in turn may necessitate measurement of different lab values or parameters, increasing complexity.
There is more scientific literature
In an echo of Moore’s law, the most interesting paper I discovered found that, dating back to 1665, the number of medical journals roughly doubles every 20 years, as shown below.
The actual number of medical journals by year (black) vs. a modeled doubling every 20 years (grey). Figure from Ghasemi et al, 2023.
However, the number of journal articles has enjoyed an even greater increase in recent years, as the paper’s Figure 2 shows below. I hypothesize that each journal can run more articles because of the shift to digital. Bytes are far cheaper than ink, so articles that simply would’ve missed the cut before can now be included in online-only supplements.
The growth in number of total articles (grey) outpaced the predicted doubling of medical journals from 2000-2020, especially from 2010 onward. Necessarily, there must be more articles per journal. Figure from Ghasemi et al, 2023.
Cancer represents an increasing share of medical literature
Finally, the most surprising paper I found notes that, from 1950 to 2016, cancer-related entries increased from 6% to over 16% of PubMed, the preeminent public database of medical literature. As the world’s population has aged, chronic diseases like cancer have grown in importance as compared to infectious diseases, especially in developed countries like the U.S.
2. Oncology workforce shortages
Our next key challenge facing the oncology community is a workforce shortage. The U.S. has 28% fewer doctors per capita than its average peer country and the gap is widening, per a 2018 report from the Peterson Center on Healthcare and the Kaiser Family Foundation.
Why?
First, the U.S. population has increased faster than in many other developed countries, so we would have needed to mint more doctors just to keep up with this baseline population growth. (As an extreme example, consider Japan, where the population has decreased for 12 straight years.) This likely understates matters in oncology, where disease burden is growing faster yet, as discussed above.
However, the number of physician training (residency) slots has barely budged for decades. This op-ed in Medical Economics is a fascinating read, but in brief: The U.S. government is the major funder of residency slots, and physician groups in the 1990s successfully lobbied for caps on these slots, fearing a physician surplus (which, in turn, could suppress wages). We physicians have ourselves to blame.
Physician dissatisfaction. Anecdotally, as a physician working in tech, I get outreach every few months from former classmates and colleagues looking to make the same transition, and the data bears out this trend: Half of doctors report burnout, per a 2022 AMA survey, 56% report high levels of job stress, and 40% of doctors are at least moderately likely to leave their jobs in the next two years.
Admittedly, the U.S. has more per-capita nurses and a faster rate of increase in nurses than peer countries, per the Peterson report. However, many nurses only work under supervision of a physician, suggesting that physician availability is the rate-limiting step. Plus, nursing is a broad field, so while the overall field may be growing quickly, the percentage of nurses with specialized training to work in a given specialty, such as oncology, or a specific function, such as clinical research, may not.
Finally, note that these workforce shortages create a positive feedback loop in which a hospital with fewer oncologists, for example, can see fewer patients and thus generates less revenue. In turn, that hospital is less able to afford additional oncologists or oncology clinical research staff.
3. Medicine is bureaucratic and change-averse
The pager was first used in hospitals in 1950, the year my parents were born. Yet, here’s a study from 2017, the year I graduated medical school, noting that 80 percent of hospital-based clinicians still used pagers.
Why is medicine so averse to change? While I have fewer peer-reviewed studies to cite a question of culture, rather than epidemiology, I do have several ideas.
Most importantly, medicine has good reason to be cautious: “Move fast and break things,” Silicon Valley’s historic ethos, simply can’t work in life-or-death situations, full stop. Many innovations ultimately don’t pan out, and a little caution is certainly warranted.
Lengthy and complicated adoption cycle: In turn, a health system operationalizes this caution by requiring multiple departments and committees to signoff on any new technology before it’s brought onboard. Again, it’s a tradeoff as these bureaucratic processes presumably weed out bad technologies. However, giving enough different groups functional veto power over a new idea also means that many promising ideas are killed off prematurely.
Vendor lock-in: Say you’re not thrilled with your electronic health record (EHR) and would like to switch to a more technologically-advanced option. The costs–not just money, but also labor, time and operational disruption–may be prohibitive and lock you into your suboptimal solution. In turn, your EHR vendor knows that, which leads to…
…Limited vendor incentive to innovate: As it’s relatively easy for consumers to switch smartphones, Apple and its competitors will spend billions of dollars in R&D annually to retain existing customers and win over new ones. My utility provider, or your EHR vendor, does not face this innovative pressure.
Limited health system incentive to innovate: I recently switched to a credit card offering a higher reward rate, because I could easily quantify the amount of money I would save, knew that these benefits would in fact accrue to me, and concluded that the hassle was therefore worth my while. By contrast, say a vendor offers a health system a technology that clearly improves a backend function. As hospital finances are beyond complex, the decision-maker may not be able to estimate the resulting savings, and further, may not be the one to solely realize the benefits. Even if it’s a good idea in theory, in practice, it may be easier for the decision-maker to default to no, especially because of…
Competing priorities: It’s a lot of work to onboard a new process or technology. If cancer centers are already short-staffed with their current workloads, as discussed above, they may simply lack the bandwidth to implement new technology, even if it would be time-saving in the long run.
In sum, the problem of antiquated technology and workflows comes with real human costs. Consider the much-hated, yet ubiquitous EHR: Primary care physicians (PCPs) with moderate or high EHR usage reported lower satisfaction than their low EHR-usage peers, per a 2014 study of over 400 PCPs and managers, and the moderate-EHR group also reported more stress than their low-usage peers. In turn, the U.S. Agency for Healthcare Research and Quality (AHRQ) recognizes EHRs as a major source of physician burnout, alongside the related challenges of time pressure and a chaotic work environment.
Major drivers of physician burnout. Figure from AHRQ.
4. Downstream effects of oncology workflow challenges
To summarize what we’re covered so far:
The oncology care and research burden is greater than ever;
There aren’t enough people to do all the work; and
Medicine is hesitant to adapt new processes or technologies that could help.
In this section, I’ll address a few effects of this mismatch between the demands facing oncology providers (section 1) and their capacity (sections 2 and 3) to meet these demands:
Costs to patients Oncology providers’ work has life-altering implications. As antiquated, time-intensive workflows divert oncology teams’ time away from patient care, patients are affected in several ways, including:
Access to care: A patient may have to wait months for an appointment, test or procedure due to staffing shortages.
Delays in care: Doctors are trained that “time is brain” in a stroke, and that “time is muscle” in a myocardial infarction (heart attack). The same logic holds in oncology. Even after the patient establishes care, administrative delays may pile up at each subsequent step of the patient’s treatment journey, despite best efforts of staff to prioritize those with more aggressive disease.
Chart review: Perhaps the provider’s EHR can’t access relevant patient notes or test results from a different health system. Even if the data is there, it might not be found, as even the most diligent oncologist must necessarily skim the records of the hundreds of patients they’ll see each month. (In contrast, generative AI could “read” and summarize most, if not all, of the patient record.)
Reactive trial staffing: Owing to these workforce and technologic limitations, most patient-trial matching is currently reactive: upon patient or provider request, clinical research staff manually perform this task. Moving to a more automated solution, like we offer at Triomics, allows for patient-trial matching to instead occur proactively: clinical research staff can now suggest potential clinical trials to providers ahead of their patients’ visits.
Costs to staff
As discussed above, a high percentage of physicians are dissatisfied with their jobs and considering quitting, and dissatisfaction with their health tech (EHRs) is a major cause. Now, imagine you’re a clinical research coordinator who also is required to use antiquated technology to solve high-stakes problems in the same field of oncology. You also have a lower salary, less organizational clout, and lower barriers to switching careers than an oncologist. Unsurprisingly then, clinical research staff burnout and turnover is indeed quite high. In turn, this staff turnover affects the job satisfaction, workload and turnover of others in the organization.
Costs to the business
Say a hypothetical cancer center could see 20% more patients were they not as understaffed, undertooled and overworked. To a first approximation, the direct opportunity cost of the center’s workflow challenges would be 20% of their current revenue. In turn, capturing this “missing” revenue may help cancer centers turn currently unprofitable business units–like registry or clinical research–profitable.
But an estimate of just direct opportunity costs likely understates matters, as it ignores the harder-to-measure effects of quality. Existing patients may not receive as many services as they should due to workflow challenges, which would impact the patient most importantly, but also the center’s revenue. In turn, these patients and their providers may be less likely to refer their colleagues, which would also affect the center’s long-term prospects.
Disadvantaged patients disproportionately bear these costs
Finally, as a physician, I’m well-equipped to advocate for my own or my loved ones’ medical needs:
I have the knowledge to proactively seek care when appropriate.
I can navigate the system to find a high-quality doctor or facility that provides this care.
I can access this care because of my insurance.
Once there, I can effectively voice my concerns.
Finally, my concerns are likely to be taken seriously.
However, the calculus is obviously different for patients with lower income, educational attainment or English fluency, to give a few examples.
A final note is that were high-quality care universal, self-advocacy would be less important, but in reality, medical error is the third-leading source of death in the U.S. It stands to reason that the workflow challenges I’ve discussed affect quality, and in turn, that disadvantaged patients are most affected.
5. How technology, and genAI, can be a partial solution
Technology is increasingly essential to the workflows of modern medicine, be it the lowly pager, the EHR or telehealth platforms. As the saying goes, you can’t boil the ocean (especially when the “ocean” of healthcare represents 17% of U.S. GDP), so at Triomics, we focus on using generative AI (genAI) to improve workflows in oncology research and care. However, even that mandate is too broad, as the diagram below of 20 genAI healthcare use cases suggests, so I have highlighted in blue the key use cases that we tackle.
Value (left diagonal) and feasibility (right diagonal) of key genAI use cases in healthcare. Triomics use cases are highlighted in blue font. Inspired by a similar chart from ValidAI.
A few additional points of context are worth calling out:
First, genAI is especially valuable in oncology because oncology is so computationally intensive. I discuss this in more detail in section #1, but consider also the growth of precision oncology, and all the associated genomic and proteomic data.
Healthy competition in the sector should ultimately benefit patients–like in my smartphone example above–and no single player will (or should) solve everything. Though the field took off just a few years ago, I’m already humbled by all the innovation in the space. My former employer Doximity lets clinicians leverage a souped-up chatGPT to, for example, generate patient instructions or insurance appeals. When I broke a bone in residency, I turned to Nuance to dictate notes, and they’ve since incorporated a suite of genAI-powered features into their software. I happened to personally interface with these two products, but in truth, this section could be 100 companies long and is growing by the month.
GenAI has significant shortcomings. It “hallucinates”, or invents facts when it is unsure of the correct answer. Given the stakes of oncology, human alignment and verification is therefore required to ensure accuracy. The computational demands are often sufficiently large that these systems can’t be hosted locally, and the offsite data transfer and storage confers additional cost and security risk. The technology is so new that the regulatory landscape is lagging, and products are often v1 releases that have yet to benefit from massive field-testing and refinement. You get the idea, but this list is also lengthy and growing.
Depth matters more than breadth. Given the technical complexity, the novelty of genAI, and the no-fail nature of the oncology space, we believe a company that solves a few problems exceptionally well will win out over time over a company that focuses instead on breadth. At Triomics, we therefore focus on building a best-in-class proprietary large language model, OncoLLM, to parse unstructured health data. We then build bespoke software that leverages our model to streamline specific tasks that cancer centers must accomplish, such as (a) patient-trial matching, or EHR data curation for (b) cancer registry reporting, or (c) integration with EDCs (electronic data capture software for clinical trial management).
Thanks for reading and I hope you found this whitepaper enjoyable and informative. Please reach out at danielmd@triomics.com if you have any further questions, or would be interested in partnering together to improve oncology workflows at your cancer center.
