Health Records Reveal Four Distinct Paths of Cancer Symptoms Over Time

Cancer rarely announces itself through a single symptom. For most patients, the disease and its treatment unleash a cluster of intertwined burdens: sleep that will not come, pain that will not fade, bodies that will not move as they once did, anxiety and depression that shadow every scan, and a fatigue so profound that researchers have given it its own name, energy deficit. A new analysis of electronic health record data from nearly fifteen thousand patients with cancer has now mapped, with unusual statistical rigor, how these six symptoms evolve over a full year, and the picture that emerges is both reassuring and sobering. Roughly two-thirds of patients follow a stable, low-severity course, but a substantial minority remain trapped in persistent high-severity trajectories or watch their symptoms steadily worsen, and the factors that predict which path a patient will take are as much social as they are medical.

The study, published in Supportive Care in Cancer, drew on the Enhanced EHR-facilitated Cancer Symptom Control trial, known as E2C2, a pragmatic cluster-randomized trial conducted across Mayo Clinic oncology practices. Between March 2019 and January 2023, more than fifty thousand adults with solid and liquid tumors were offered symptom screeners through the electronic health record in connection with their oncology visits. From this cohort, the researchers identified 14,590 patients who had completed at least three symptom assessments within a twelve-month window, a threshold chosen to sharpen the precision of individual trajectory estimates. The six symptoms under scrutiny, abbreviated SPPADE, were sleep disturbance, pain, physical function impairment, anxiety, depression, and energy deficit or fatigue. Each was rated by patients on an eleven-point numeric scale from zero, meaning no symptom, to ten, meaning as bad as one can imagine, capturing the preceding week of experience.

What makes this analysis methodologically distinctive is its data source. Most trajectory studies in oncology rely on panel data, in which research staff assess every participant at predetermined intervals. Such designs are expensive, small, and often restricted to a single cancer type or treatment stage. Here, the researchers instead mined routinely collected clinical data, in which assessments occur at irregular intervals determined by visit schedules and patient willingness to complete optional surveys. Latent class growth analysis, a statistical technique designed to identify unobserved subgroups that follow similar patterns of change over time, proved robust to this irregularity. The models included three growth factors, an intercept, a linear slope, and a quadratic term, allowing for non-linear change, and were fitted with two thousand random starts to ensure stable solutions.

Selecting the optimal number of trajectory classes required a careful balancing act across five fit indices. The Bayesian Information Criterion kept declining as classes were added, offering no clear elbow, so the researchers leaned on entropy, the Lo-Mendell-Rubin likelihood ratio test, average posterior probability, and minimum class size thresholds. Five-class solutions repeatedly produced subgroups smaller than five percent of the sample, undermining their reliability. For four of the seven symptom scores examined, including the composite burden score, the four-class solution won on multiple criteria simultaneously. Even where fit statistics pointed elsewhere, clinical interpretability prevailed: for pain, a three-class solution would have erased the worsening trajectory, which affected nearly nine percent of patients and was judged too clinically important to discard.

The result was a remarkably consistent architecture across all six symptoms. On average, 65.5 percent of patients occupied a low-severity stable trajectory, 15.5 percent a high-severity stable trajectory, 10.7 percent an improving trajectory, and 8.4 percent a worsening trajectory. These proportions held with striking uniformity whether the outcome was pain, anxiety, fatigue, or the composite SPPADE score summing all six symptoms from zero to sixty. The authors interpret this consistency as evidence of common underlying dynamics governing symptom stability and change, suggesting that whatever drives a patient toward persistent suffering or toward recovery operates across symptom types rather than being unique to each one.

Just as consequential is what predicted trajectory membership. Using standardized mean differences to screen for imbalance and then multivariable multinomial logistic regression to isolate independent associations, the researchers found that employment status carried the strongest signal, followed by educational attainment, marital status, metastatic disease, and age. Being employed, married or partnered, better educated, older than sixty-five, and free of metastatic cancer each independently increased the likelihood of following the favorable low-stable course. Conversely, patients who were disabled or unemployed, unpartnered, or living with metastatic disease faced significantly elevated relative risks of high, worsening, or changing symptom trajectories across multiple symptoms. Middle-aged patients between forty and sixty-four showed, for example, a 1.5-fold higher relative risk of a worsening physical function trajectory compared with patients under forty, relative to the low-stable comparator.

Perhaps the most surprising finding was what did not matter. Cancer site, so often the organizing principle of oncology research and care, was generally unrelated to symptom trajectory, with standardized mean differences ranging only from 0.06 to 0.15. The exceptions were instructive: fatigue was less likely to improve across nearly all cancer sites compared with remaining at a low stable level, and persistently poor sleep was less likely at several sites. The authors caution, however, that the predominantly White, well-educated sample limits generalizability, and that the small subgroups of African-American and Hispanic patients, each numbering only 132, showed elevated relative risks across multiple symptoms that demand cautious interpretation and further study in more diverse populations.

The clinical implications are immediate. Oncology practices that already administer patient-reported symptom screeners, as E2C2 did through both in-clinic and remote monthly assessments, possess a continuously updating dataset that can stratify patients without any additional data collection. Patients in the low-stable majority might be spared intensive surveillance and offered self-management resources, freeing clinical capacity for those in the high-stable or worsening groups who need collaborative care interventions. The E2C2 trial itself demonstrated that such interventions work, reducing symptom burden while cutting emergency department visits, hospitalizations, and intensive care admissions. Because single-item numeric rating scales are public domain, responsive to change, and preferred by patients over visual analogue scales, the infrastructure for this kind of stratified care is inexpensive and scalable.

The study also opens a research frontier. The authors propose investigating predictors of improvement among severely symptomatic patients and of worsening among those who start low, tracking treatment-related trajectories through chemotherapy, radiation, and surgery, incorporating laboratory values and medication data, and developing finer markers of cancer stage beyond metastasis. Machine learning approaches that predict future symptom trajectories from prior EHR data are already emerging, and this study provides the validated trajectory framework those models will need to predict against. The researchers are candid about limitations: trajectories began at first screener completion rather than diagnosis, intervention effects could not be disentangled, and the analysis identifies associations rather than validating a prospective prediction model. Still, the temporal logic, with baseline characteristics preceding subsequent trajectories, makes reverse causation unlikely.

For a field that has long treated cancer symptoms as static problems to be assessed at single time points, this study delivers a conceptual shift: symptoms have histories, and those histories follow recognizable patterns. Two-thirds of patients may weather their illness with stable, manageable symptoms, but the one-third who suffer persistently or deteriorate can now be identified earlier, using nothing more than the data their health system already collects. The social gradient embedded in the findings, with employment, education, and partnership shaping symptom outcomes as powerfully as tumor biology, is a reminder that supportive cancer care is not only about the cancer. It is about the whole life that the disease interrupts, and about building systems nimble enough to notice when that life begins to buckle under the weight of unrelieved suffering.

Subject of Research: Longitudinal trajectories of sleep, pain, physical function, anxiety, depression, and fatigue symptoms in patients with cancer identified through electronic health record data

Article Title: Identifying the trajectory of SPPADE symptoms in patients with cancer using electronic health record data

Article References: Identifying the trajectory of SPPADE symptoms in patients with cancer using electronic health record data. (n.d.). https://doi.org/10.1007/s00520-026-11234-4

Image Credits: AI Generated

DOI: 10.1007/s00520-026-11234-4

Keywords: cancer symptoms, symptom trajectories, electronic health records, patient-reported outcomes, latent class growth analysis, SPPADE symptoms, oncology, supportive care, fatigue, pain, health informatics, symptom management