With Cancer Prognoses, Are We Relying Too Much on Biomarkers and Algorithms?

Ravi B. Parikh, MD, MPP


March 01, 2019

Risk Assessment in Oncology—A Tale of Two Oncologists

The pathology report hit me like a pile of bricks: "PD-L1 high (CPS 90%)."

In most cases of newly diagnosed non–small cell lung cancer, I would have been happy to see this. It would mean that my patient could receive checkpoint inhibitor therapy alone—without chemotherapy—in the first line with reasonable hope of a durable response. We could spare the patient chemotherapy, at least for now.

But this patient was different. In his late sixties, he came to the hospital with weeks of worsening shortness of breath and headache. Recently he felt a dull pain in his abdomen that wouldn't go away. Imaging revealed what we feared: a dominant lung mass, with countless metastases throughout his lungs, liver, and bone, along with a large adrenal mass. He also had a sizable frontal brain tumor. And as a longtime smoker with multiple vascular issues, he was far from healthy coming into this.

Why not give the patient a chance?

His pathology had come back suggesting adenocarcinoma, a primary lung cancer. Given how rapidly his cancer was progressing and how symptomatic he was, he needed chemotherapy—and fast. Unfortunately, he was only getting weaker and more short of breath.

After several days in the hospital, our team approached his wife and the patient with sobering news: "Things aren't getting better. Your body isn't letting us give the therapy needed to treat your cancer."

After some discussion, the family agreed with us and were thinking seriously about enrolling in hospice.

As we stood up, the wife asked, "Is there a chance he could receive therapy?"

We fidgeted a bit, wondering whether we should even mention the fact that he could still be eligible for less toxic therapies, depending on genetic or molecular testing.

"Let's see what the pathology shows," we replied.

A Whole New World of Risk Assessment

The next day, the PD-L1 results came back, throwing a monkey wrench into the conversation about hospice that I'd had with the family.

I spoke to two oncologists about the case. One agreed that pursuing hospice was still the right course of action. "He probably won't live long enough to benefit from the drug," he said, noting that immunotherapy can take weeks to have a response.

The second oncologist, however, was more optimistic. Granted, he believed that the prognosis was still poor. But if the patient could safely receive immunotherapy, why not give him a chance? "There's a reasonable chance that it could prolong his life, and maybe his quality of life."

I imagine that these days, a lot of us might agree with the second oncologist. Next-generation sequencing and biomarker testing has skyrocketed[1] in diseases like non–small cell lung cancer. While less than 5% of patients' treatments are influenced by this testing,[2] there is considerable excitement that this number will increase as new therapies become available.

Oncologists have been notoriously poor at prognosis.

As our ability to determine genetic mutations and molecular-expression profiles of cancers has improved, we are using these biomarkers more and more to determine patients' risk for progression or death. Risk stratification in acute myeloid leukemia[3] and node-negative, hormone-positive breast cancer[4] is now nearly entirely based on genetic or molecular characteristics. And we rely on this risk stratification to determine treatment.

Granted, biomarker-based risk assessment can be a good thing. Previous nomograms and rules to determine risk were based entirely on clinical characteristics like labs, pathology, and performance status (PS). The D'Amico risk score for localized prostate cancer is a good example of a previous generation of risk assessment tools. Based on a limited number of variables—serum prostate-specific antigen (PSA) levels, clinical stage, and Gleason score—the D'Amico risk score has been used for over two decades to determine how aggressively to treat localized prostate cancer and to screen patients into clinical trials.

On average, doctors overestimated survival by a factor of five.

But often these characteristics can vary significantly and are extremely subjective. Pathologists agree on the Gleason score less than 60% of the time,[5] a troubling statistic. And oncologists disagree 35% of the time[6] when rating a patient's Eastern Cooperative Oncology Group (ECOG) PS. Yet, PS is a key input for many prognostic indicators in oncology.

The pitfalls of clinical risk assessment may explain why oncologists have been notoriously poor at prognosis. In perhaps the most widely cited study demonstrating this, published in 2003 in the British Medical Journal,[7] nearly 350 physicians were asked to predict the life expectancy of patients at the time they entered hospice. These patients were already at a high risk for death, and it should have been relatively easier to predict life expectancy in this cohort.

However, only 20% of predictions were considered accurate—that is, the oncologist's predicted survival fell within a range that was plus or minus 33% of the actual survival. The vast majority of predictions (63%) were overoptimistic, compared with 17% which were overly pessimistic. On average, doctors overestimated survival by a factor of five.

But Is Biomarker-Based Risk Assessment Better?

Certainly, biomarker-based risk assessment offers advantages compared with this traditional paradigm. Genes and molecules are objective and theoretically expressed throughout a tumor specimen (although issues with heterogeneity still exist). Furthermore, you can do something about a tumor with poor-risk biomarkers: recognizing that HER2-expressing tumors conferred a poor prognosis, drug makers developed trastuzumab. Now, in a modern era of HER2-directed agents, HER2 expression is not nearly as dire as it used to be. It's much more difficult to develop therapies to address clinical risk factors like PS.

I really had no evidence to rely on.

The problem occurs when we assume that biomarkers, genetics, and molecular-expression profiles are surrogates for clinical risk factors, such that we ignore those clinical factors entirely. I recently saw a young woman in clinic with ALK-rearranged stage IV lung cancer. She had been doing well on ALK-targeted therapy for a couple of years. But she had recently developed an early-stage hormone-positive breast cancer. When she asked about her risk for breast cancer recurrence, I discussed the need to wait for her tumor genomic testing results to determine her risk and need for chemotherapy.

But when I stepped out to discuss her case with my attending, he stopped me quickly when I began rambling about the Oncotype DX test.

"Ravi, she has stage IV lung cancer. Let's not lose sight about what's going to kill her."

He was right, of course. Her stage IV lung cancer was the major determinant of her prognosis. No matter what her genetic risk score was, we would not be offering adjuvant chemotherapy to improve her risk for 10-year survival by a few percentage points. Quality of life was key here. And as my mentor reminded me, the practice-changing trials that we often cite routinely excluded patients with poor PS and secondary cancers. I really had no evidence to rely on.

It was a sobering lesson in what really matters when we talk about prognosis in oncology.

But as my attending pointed out to me later, there are many oncologists who would have offered her therapy if her genomic risk score was high. Indeed, her breast surgeon had already sent the test out.

A Path Forward

To be fair, newer risk assessment tools have tried to integrate these two parallel schools of thought—biomarker and clinical risk assessment. The revised international staging system for myeloma (R-ISS) now incorporates the presence of high-risk chromosomal abnormalities into its predictive tool, based primarily on clinical risk factors. A similar effort in metastatic renal cell carcinoma was presented as a poster[8] at this year's ASCO Genitourinary Cancers Symposium. Other risk tools will probably increasingly incorporate predictive and prognostic biomarkers. And machine learning algorithms, which are able to account for thousands of genetic and clinical predictors, will only make this easier.

But as we increasingly have this information in oncology, we must remember that the most important prognostic variable is the "eye test"—how the patient in front of us looks. It's exceedingly rare that a next-generation sequencing or immunohistochemical test report can reveal a potential treatment that will meaningfully or quickly improve a patient's clinical status.

Prognosis and risk assessment are still the domains of the clinician, not the algorithm.


The next day, I made my way into the hospital to tell the patient and his family about the results of his PD-L1 testing. I wasn't quite sure how I was going to talk about future therapy. But a group of nurses were blocking the entrance to the patient's room. He had developed acute shortness of breath overnight and a rapid response had been called. Imaging showed a large pleural effusion—likely malignant—that had developed nearly overnight. The clinical team was deciding whether to transfer the patient to the intensive care unit.

His body had made the decision for us. No matter what his PD-L1 expression was, he was in no state to receive treatment.

He and his family, after a long discussion, decided against the ICU transfer, and a few days later he was enrolled in an inpatient hospice program. The patient and his medical team—not his genetics or molecular profile—had determined his care, as should still be the case for all patients with cancer.

Follow Medscape on Facebook, Twitter, Instagram, and YouTube


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.