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

Abstract

Purpose

The longitudinal course (trajectory) of symptoms in patients with cancer has been understudied. Sleep disturbance, pain, physical function impairment, anxiety, depression, and energy deficit/fatigue (SPPADE) are particularly common in patients with cancer. This study characterizes the trajectory patterns of SPPADE symptoms in a large and diverse sample of patients with cancer and identifies factors associated with distinct trajectories.

Methods

To identify the optimal number and functional forms of the symptom trajectories, latent class growth analysis was conducted. Electronic health record (EHR) data were analyzed from 14,590 patients in the Enhanced EHR-facilitated Cancer Symptom Control (E2C2) trial cohort with at least 3 symptom assessments within a 12-month period. Each SPPADE symptom was rated on a 0–10 numeric rating scale. Multivariable multinomial logistic regression was used to determine factors associated with membership in the symptom trajectory subgroups.

Results

Four distinct trajectory patterns were similar across symptoms. The average sample distribution of membership in a trajectory class across the 6 symptoms was 65.5% for low severity stable, 15.5% for high severity stable, 10.7% for improving, and 8.4% for worsening. Being employed, married/partnered, better educated, older, and not having metastatic disease were independently associated with membership in the low stable trajectory of SPPADE symptoms. Cancer site generally did not influence symptom trajectory class membership.

Conclusion

For oncology practices that routinely use patient-reported outcomes to assess symptoms, EHR data can be a source for characterizing symptom trajectories. Knowledge of factors associated with trajectory class membership could be useful in clinical research and for tailoring evidence-based symptom management strategies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s00520-026-11234-4.

Introduction

Symptoms are a major consequence of cancer and its treatment and adversely affect quality of life, work productivity, and health care utilization [1]. Sleep disturbance, pain, physical function impairment, anxiety, depression, and energy deficit/fatigue (SPPADE) are a particularly common set of symptoms that have high rates of co-occurrence, undertreatment, and persistence during and following cancer treatment [2]. The Enhanced EHR-facilitated Cancer Symptom Control (E2C2) trial is a stepped wedge, cohort cluster randomized, pragmatic clinical trial designed to compare the effectiveness of routine symptom surveillance with surveillance and collaborative care model-based symptom management targeting six SPPADE symptoms [2, 3]. The E2C2 intervention was effective in reducing symptom burden as well as decreasing emergency department visits, hospitalizations, and intensive care unit admissions [4].

Research on the longitudinal course of cancer-related symptoms is nascent and has predominantly involved relatively small samples or has focused on a single symptom, one site of cancer, or a specific treatment stage (e.g., active treatment, palliation, survivorship or end-of-life) [5–16]. Most trajectory analyses use panel data (i.e., outcomes assessed by research personnel on all participants at pre-defined intervals). Few use patient data from the electronic health record (EHR) which comprises a varying number of symptom assessments collected at irregular intervals. Trajectory analysis of clinical data has the potential of expanding our understanding of the longitudinal course of cancer-related symptoms in much larger numbers of patients in real-world clinical settings.

In this paper, we analyze longitudinal EHR data from the E2C2 trial to characterize the trajectories of SPPADE symptoms in a large and diverse sample. Our objectives are twofold. First, we use latent class growth analysis to identify the number and frequency of discrete trajectories for each of 6 SPPADE symptoms as well as a composite score of symptom burden. Second, we use multinomial logistic regression to examine which demographic and clinical characteristics are independently associated with symptom trajectory. Determining what influences the longitudinal course of cancer symptoms may help identify prevention and treatment targets as well as tailor or intensify symptom management strategies.

Methods

Study sample

The E2C2 trial population included adult patients (≥18 years) with malignant solid and liquid tumors. All cancer stages, cancer sites, and treatment phases were included except for patients enrolled in hospice or with acute leukemias. An EHR-based algorithm ascertained eligibility by requiring patients to have a provider visit in a participating hematology/oncology department linked to an ICD-10 cancer diagnosis. Patients were followed continuously for outcome assessment until the end of the trial, a maximum of 47 months, excepting the cases of death and exit from the health system.

From March 2019 to January 2023, 50,207 patients participated in this pragmatic trial and were offered SPPADE symptom screeners through the EHR in association with oncology visits that occurred at least 14 days apart. Of the full trial cohort, 38,644 completed one or more symptom screeners. Among these, 14,590 provided symptom scores at least three times during the first year of surveillance and comprised the analytic cohort for this trajectory modeling. We restricted the sample to those with ≥3 timepoints to improve precision of individual-level trajectory class estimates and enhance the reliability of trajectory class assignment [17]. We selected a 1-year window to standardize follow-up duration among individuals and to minimize confounding from changes in cancer course or treatment. Because patients could complete symptom screeners during the control period, intervention period, or both periods, data was included for eligible patients across all periods.

The E2C2 study (clinicaltrials.gov NCT03892967; registered 3/25/2019) received institutional review board approval with waiver of informed consent because symptom monitoring was considered a standard of care and the trial’s automated clinical responses to symptom reporting are guideline-supported and an evidence-based best practice [18]. Additional details of the E2C2 trial are described elsewhere [2, 3].

Symptom assessment

A 6-item symptom screener was administered to patients prior to, during, or soon after outpatient encounters that occurred at least 2 weeks apart in participating Mayo Clinic oncology clinics. During intervention periods, participants additionally had the option of completing monthly remote assessments. The screener included an 11-point numeric rating scale (NRS) for each of the six SPPADE symptoms wherein patients described their symptom levels in the past week from 0 (none) to 10 (as bad as you can imagine). NRSs have been used extensively among patients with cancer of all types and stages [19, 20] and have also been one of the instruments most often used to measure symptom ratings in NCI cancer clinical trials [21]. Additionally, NRS are more responsive than either Likert or verbal rating scales [22] and are preferred by patients over visual analogue scales [23]. We analyzed responses for each symptom separately, as well as a composite score created by summing across all size symptoms (range 0–60).

Other variables

Sociodemographic variables included age, sex, race, ethnicity, education, employment status, marital status, payor/health insurer and rural–urban commuting area (RUCA) codes. These were abstracted from the Mayo Clinic EHR, Epic, which was shared across all participating sites. Patient portal use was defined as the patient having accessed the Epic health care portal at least once in the 12months prior to first E2C2 contact [24]. Cancer site and metastatic status were determined from the EHR as previously described [25].

Statistical analysis

Determining optimal number of symptom trajectories

Latent class growth analysis (LCGA) [17, 26] is a statistical methodology used to identify unobserved subgroups (latent groups) within a population that follow similar patterns of change over time. LCGA was used to classify patients into varying numbers of symptom trajectories ranging from 2 to 6. Each model included three growth factors: intercept, linear slope, and quadratic term, which allowed for non-linear change over time. Within-class quadratic variance was fixed to zero due to convergence issues and negative variance estimates, which occur when quadratic effects are small or highly correlated with linear terms. A set of 2000 random starts was used, retaining the best 500 for final optimization. Models were estimated using maximum likelihood with robust standard errors under the assumption of missing at random.

Five fit indices were examined to inform selection of the optimal number of latent trajectories [17, 26]. The Bayesian Information Criterion (BIC) is one important fit index for which lower values indicate superior fit. The BIC often continues to decrease for each additional class added (e.g., there is not a global minimum) in which case an elbow in the BIC curve (i.e.., small decreases in the BIC for each additional latent class) may help determine the number of classes. The Lo, Mendell, and Rubin likelihood ratio test (LMR) compares the fit between two neighboring class models wherein a non-significant p-value for a k class solution supports the k−1 class solution. Entropy is an omnibus index where values >0.80 indicate “good” classification of individual cases into classes. Average posterior probability (AvePP) indicates how well a given model classifies individuals into their most likely latent trajectory class. An AvePP > 0.70 indicate well-separated classes. A minimum class size criterion recommends that each class included in a model represents at least 5–8% of the overall sample. There is no consensus on a single “best” fit index, which is why a set of indices are jointly used to assess statistical fit and then integrated with theory, clinical interpretability, and utility [27].

Determining factors associated with symptom trajectory

Associations between demographic and clinical variables and latent trajectory group membership were explored in a multistep process. First, we assessed imbalance in sociodemographic and clinical characteristics between trajectory groups by examining the standardized mean difference (SMD). SMD is calculated as the difference in means or proportions divided by the standard deviation and provides a scale-free measure of difference between groups. To calculate the SMD for variables with more than 2 categories, we used the method of Yang and Dalton as implemented in Stata’s -stddiff- package [28]. To conservatively estimate the SMD for a variable across 4 trajectory categories, we report the maximum of all pairwise comparisons. SMDs are less sensitive to sample size than p-values, making them preferable for variable screening [29] when no a priori hypotheses exist for which trajectory group would have higher or lower values; variables with an absolute SMD ≥ 0.20 [28] were considered to show potentially meaningful imbalance across trajectory groups.

Second, bivariate multinomial models were then estimated for each variable across all symptom trajectory sets and the composite outcome. Variables that were significant at P < 0.05 in at least one bivariable model were then included in multivariable multinomial models for all symptoms. These final models represent the independent associations of each baseline variable and its subcategories with specific symptom trajectory groups. Multinomial model effects are reported as relative risks (RRs). Variables for which the 95% confidence intervals did not include 1.0 were considered factors that have a potential independent association with symptom trajectory. In summary, an independent association between a variable and symptom trajectory was indicated by either an SRM ≥ 0.20 for that variable or by a RR for which the 95% CI does not include 1.0.

The latent trajectory analyses were performed in Mplus version 8.11. All other analyses were performed in SAS version 9.4 (SAS Institute) and Stata version 18.0 (StataCorp, College Station TX). This trial was approved by the Mayo Clinic IRB.

Results

Study sample

Characteristics of the 14,590 E2C2 participants in the trajectory analysis sample are summarized in Table 1. Most were in the 40–64 (45%) or ≥65 (49%) age groups. The sample was 57% female and predominantly White and of non-Hispanic ethnicity. Participants had relatively high levels of educational attainment and were mostly either employed or retired. About a fourth (27%) of the sample resided in a small town or rural area. The cancer was metastatic in 46%. The most common cancer sites were gastrointestinal (26%), breast (24%), lung (15%), genitourinary (14%), and skin (12%). Two or more cancer sites were recorded in the EHR for 42% of the sample.

Table 1.

Characteristics of E2C2 participants included in trajectory analysis (N = 14,590)

Characteristic	N	(%)
Age
<40	926	(6.4)
40–64	6593	(45.2)
≥65	7071	(48.5)
Sex
Female	8378	(57.4)
Male	6212	(42.6)
Ethnicity
Not Hispanic or Latino	14,409	(98.8)
Hispanic or Latino	132	(0.9)
Unknown/Not Reported Ethnicity	49	(0.3)
Race
White	13,935	(95.5)
African-American	132	(0.9)
American Indian/Alaska Native	63	(0.4)
Asian/Pacific Islander	251	(1.7)
Other/Unknown	209	(1.4)
Marital status
Married/Partnered	10,526	(72.2)
Divorced/Separated	1300	(8.9)
Widowed	1177	(8.1)
Single	1564	(10.7)
Unknown	23	(0.2)
Employment
Employed	5985	(41.0)
Retired	6934	(47.5)
Not Employed/Student/Military	1108	(7.6)
Disabled	553	(3.8)
Unknown	10	(0.1)
Health Insurance Payor
Government	8894	(61.0)
Non-government	5663	(38.8)
Unknown	33	(0.2)
Education
Less than high school	3080	(21.1)
High school	477	(3.3)
Some college or Associate degree	4139	(28.4)
Bachelor’s degree	2986	(20.5)
Master’s or Doctoral degree	2149	(14.7)
Unknown	1759	(12.1)
RUCA Code
Urban (1–3)	8058	(55.2)
Micropolitan (4–6)	2554	(17.5)
Small town (7–9)	2086	(14.3)
Rural (10)	1834	(12.6)
Unknown	58	(0.4)
Metastatic disease status
Not metastatic	4942	(33.9)
Metastatic	6708	(46.0)
Uncertain	2940	(20.2)
Cancer site
Breast	3503	(24.0)
Endocrine	909	(6.2)
GI	3804	(26.1)
GU	1998	(13.7)
Gyn	1272	(8.7)
Head and neck	872	(6.0)
Heme	1428	(9.8)
Lung	2205	(15.1)
Melanoma	678	(4.7)
Nervous system	1256	(8.6)
Sarcoma	913	(6.3)
Skin	1789	(12.3)
Multiple (≥2 categories present in EHR)	6075	(41.6)
Uncertain	38	(0.3)
Portal use
In 365 days before enrollment	10,972	(75.2)
During entire study period	13,742	(94.2)

 Optimal number of trajectory classes

Table 2 summarizes the fit statistics for the number of trajectory classes ranging from 1 to 6 for each of the 6 SPPADE symptoms as well as the composite score. Of note, all 5-class solutions had one class <730 (i.e., 5% of sample) which provides further support for the selection of the 4-class solution. Because BIC declined with each additional class added and there was no clear “elbow” with BIC across models, the highest entropy score and LMR test results were used to determine the optimal number of classes for each symptom score. For 4 of the 7 symptom scores (anxiety, depression, fatigue, sleep disturbance) the 4-class solution had the highest entropy score as well as an LMR p-value < 0.05. For pain, the highest entropy score supported a 4-class solution whereas a nonsignificant p-value for the LMR favored a 3-class solution. However, the 3-class solution left out the worsening trajectory which constituted 8.9% the sample for pain and was a clinically important trajectory. For physical function, the 6-class solution had the highest entropy and a nonsignificant p-value for the LMR favored a 5-class solution. However, the smallest group (worsening trajectory in the 4-class solution had 8.2% of the sample, whereas the 5- and 6-class solutions each had a class representing less than 5% of the sample. For the SPPADE composite score reflecting total symptom burden, the 5-class solution had the highest entropy but only slightly greater than the 4-class solution (0.766 vs. 0.762). Whereas the nonsignificant LMR p-value for the 4-class model favored a 3-class model, the 4-class model had the highest average posterior probability. Finally, the 4-class solution (reflecting high stable, low stable, worsening and improving trajectories) for symptoms is clinically meaningful. For these reasons, the 4-class trajectory solution was considered optimal and was used to examine the association of trajectory class membership and selected demographic and baseline clinical variables.

Table 2.

Comparison of model fit indices for the latent class growth models

Number of classes	BICb	Entropyc	Average posterior probabilityd	LMR likelihood ratio teste
Anxiety
1	219,436.359	N/A	1.000	N/A
2	216,502.736	0.813	0.893	<0.0001
3	215,013.986	0.801	0.866	<0.0001
4	213,521.296	0.844	0.861	0.0079
5	212,489.884	0.816	0.823	<0.0001
6	211,918.270	0.818	0.803	0.0046
Depression
1	218,196.431	N/A	1.000	N/A
2	215,023.894	0.817	0.920	<0.0001
3	213,316.685	0.829	0.882	<0.0001
4	211,462.637	0.860	0.872	<0.0001
5	210,364.755	0.842	0.843	0.0001
6	209,761.901	0.834	0.821	0.0264
Fatigue
1	232,460.114	N/A	1.000	N/A
2	231,186.591	0.648	0.877	<0.0001
3	230,701.323	0.692	0.816	<0.0001
4	229,676.365	0.727	0.819	<.0001
5	229,234.403	0.717	0.770	0.0005
6	228,876.259	0.720	0.748	0.2198
Pain
1	226,050.145	N/A	1.000	N/A
2	222,884.497	0.840	0.938	<0.0001
3	221,170.539	0.851	0.880	<0.0001
4	219,148.249	0.857	0.872	0.1305
5	217,906.506	0.839	0.847	<0.0001
6	217,349.918	0.827	0.823	<0.0001
Physical dysfunction
1	232,585.864	N/A	1.000	N/A
2	230,152.153	0.781	0.932	<0.0001
3	229,935.259	0.755	0.844	<0.0001
4	228,684.963	0.770	0.844	0.0133
5	227,948.329	0.761	0.808	0.0002
6	227,544.849	0.784	0.816	0.1816
Sleep disturbance
1	233,679.809	N/A	1.000	N/A
2	231,640.174	0.733	0.902	<0.0001
3	230,848.388	0.744	0.830	<0.0001
4	229,280.69	0.787	0.841	<0.0001
5	228,607.225	0.767	0.803	0.0041
6	228,081.767	0.758	0.785	0.0049
SPPADE composite
1	391,205.694	N/A	1.000	N/A
2	389,543.114	0.722	0.892	<0.0001
3	388,701.487	0.754	0.791	0.0001
4	387,709.969	0.762	0.815	0.1343
5	386,833.399	0.766	0.795	<0.0001
6	386,503.649	0.766	0.756	0.0266

^aBayesian Information Criterion (BIC), lower BIC indicates better fit

^bEntropy ranges from 0 to 1 with higher values indicating more accurate classification into latent classes

^cAverage of the probabilities that individuals are classified into a specific latent class

^dLo, Mendell, and Rubin likelihood ratio test. A significant result indicates a better fit for the tested model (K) compared to the model with one fewer classes (K-1). NA = not applicable

SPPADE symptom trajectories

Figure 1 shows the 4 trajectory classes of the SPPADE composite score, and Fig. 2 shows the 4 trajectory classes for each of the 6 individual symptoms. Across all trajectory models, trajectories we named as high stable symptoms, low stable symptoms, improving symptoms, and worsening symptoms are seen. Table 3 shows the sample proportions of patients represented by each trajectory group. Notably, the trajectory class proportions are relatively comparable regardless of symptom type, with the mean (range) sample proportions across all 6 symptom scores being 65.5% (56.7 to 72.6) for stable low, 15.5% (9.6 to 22.5) for stable high, 10.7% (8.5 to 12.1) for improving, and 8.4% (7.2 to 10.7) for worsening. Median values were close to the means. The distribution was similar for the SPPADE composite score, with slightly higher percentage (about 5%) of the sample classified into the low stable trajectory group, compared to either improving or worsening.

Fig. 1.

Symptom trajectories for the SPPADE composite score. The latter ranges from 0 to 60 as it is the sum of the 6 individual SPPADE 0–10 scores. The 4 trajectories are high stable (red line), low stable (green line), improving (blue line), and worsening (yellow line)

Fig. 2.

Trajectories for the 6 individual SPPADE symptom scores

Table 3.

Frequency distribution of SPPADE symptom trajectories (n = 14,590)

Symptom	Low stable		High stable		Improving		Worsening
	%	(N)	%	(N)	%	(N)	%	(N)
Depression	72.6	(10,596)	9.6	(1400)	10.4	(1523)	7.3	(1071)
Anxiety	71.4	(10,410)	10.3	(1495)	11.2	(1630)	7.2	(1055)
Pain	70.6	(10,297)	12.7	(1854)	8.5	(1240)	8.2	(1199)
Sleep	63.0	(9195)	16.2	(2359)	12.1	(1766)	8.7	(1270)
Fatigue	58.6	(8551)	21.4	(3115)	11.7	(1713)	8.3	(1211)
Physical Impairment	56.7	(8271)	22.5	(3281)	10.1	(1471)	10.7	(1567)
Mean of 6 symptoms	65.5		15.5		10.7		8.4
Median of 6 symptoms	66.8		14.5		10.8		8.3
SPPADE Composite	70.7	(10,318)	15.7	(2297)	7.1	(1041)	6.4	(934)

Factors independently associated with symptom trajectory

Table 4 summarizes the independent association between patient characteristics and the symptom trajectories, adjusted for other covariates. RUCA codes were not significant and thus not included as a covariate in the models. The five variables most strongly associated with differing symptom trajectories (i.e., SMDs greater than or close to 0.20) included employment status (SMD = 0.39), educational attainment (SMD = 0.24), marital status (SMD = 0.21), presence of metastatic disease (SMD = 0.19), and age (SMD = 0.19). The relative risks for membership in the high or changing (worsening or improving) symptom trajectories (compared to the low stable trajectory) were significantly greater across multiple symptoms in patients who were disabled or not employed, not married or partnered, or had metastatic disease, and lower in patients with more education or who were older than 65 years. Thus, being employed, married/partnered, better educated, older, and not having metastatic disease were independently associated with membership in the low stable trajectory of SPPADE symptoms. Notably, cancer site was not associated with symptom trajectory patterns, with SMDs ranging from 0.06 to 0.15. More detailed data on SMDs by individual symptoms are provided in the Supplemental Table.

Table 4.

Adjusted relative risk for the association of variables with symptom trajectory from multivariable multinomial modeling^a

Characteristic	N	SMDb	Worsening							High stable							Improving
			T	S	PF	P	A	D	E	T	S	PF	P	A	D	E	T	S	PF	P	A	D	E
Age		.192
<40	926		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
40–64	6593		1.4	1.1	1.5	1.1	0.7	0.8	1.1	0.9	1.0	1.2	1.1	0.6	0.7	1.0	0.9	1.0	0.8	1.1	1.0	0.9	0.9
≥65	7071		1.2	0.9	1.5	0.8	0.5	0.7	1.0	0.6	0.7	1.2	0.7	0.3	0.4	0.8	0.6	0.7	0.8	0.9	0.7	0.6	0.7
Sex		.119
Female	8378		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
Male	6212		1.0	1.0	1.2	0.8	0.9	0.9	1.0	0.8	0.9	1.0	0.8	0.6	0.7	0.9	0.7	0.9	1.0	0.8	0.6	0.7	0.9
Ethnicity		.068
Not Hispanic or Latino	14,409		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
Hispanic or Latino	132		2.1	2.1	1.3	1.4	1.5	2.1	2.2	1.5	2.0	1.8	1.0	1.2	1.3	1.6	1.5	1.4	1.4	1.3	1.7	1.4	1.0
Race		.112
White	13,935		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
African-American	132		2.7	1.9	2.0	2.8	2.6	2.0	2.5	1.5	1.4	1.5	2.2	1.3	1.6	1.6	2.6	1.7	1.0	2.2	1.9	1.9	1.6
Am. Indian/Alaska Native	63		1.2	0.7	1.9	1.1	1.2	1.1	0.7	2.1	1.9	2.0	1.9	1.9	1.3	1.5	0.5	1.5	0.6	1.0	1.4	0.9	0.1
Asian/Pacific Islander	251		1.4	1.1	1.3	0.7	1.0	1.7	0.8	1.3	1.0	1.1	1.2	0.9	1.0	0.8	1.4	0.9	1.3	1.4	1.6	1.7	0.8
Marital status		.206
Married/Partnered	10,526		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
Divorced/Separated	1300		1.3	1.5	1.3	1.4	1.2	1.5	1.4	1.5	1.4	1.4	1.6	1.6	1.3	1.3	1.2	1.2	1.3	1.3	1.0	1.1	1.2
Widowed	1177		1.3	1.2	1.1	1.1	1.0	1.2	1.2	1.0	1.0	1.2	1.1	1.0	1.1	1.1	0.9	1.0	0.8	0.9	0.9	0.9	1.0
Single	1564		1.1	1.2	1.2	1.2	1.0	1.1	1.3	1.3	1.3	1.2	1.3	1.4	1.4	1.2	1.2	1.2	1.3	1.2	1.1	1.1	1.2
Employment status		.388
Employed	5985		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
Retired	6934		1.1	1.1	1.2	1.1	1.2	1.3	1.2	1.2	0.9	1.4	1.1	1.1	1.2	1.1	1.5	1.2	1.2	1.1	1.4	1.2	1.2
Not Employed/Stud/Mil	1108		1.5	1.4	1.8	1.5	1.6	1.7	1.5	2.1	1.7	1.8	1.9	1.8	2.0	1.7	1.5	1.2	1.3	1.4	1.3	1.4	1.1
Disabled	553		2.0	2.1	2.3	1.9	2.4	2.3	2.0	3.9	2.8	4.3	3.6	2.8	3.2	3.0	1.8	1.3	1.5	1.4	1.5	1.5	1.6
Educational attainment		.241
Less than high school	3080		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
High school	477		1.0	1.0	1.0	0.8	1.2	1.2	1.2	1.3	1.2	1.0	1.4	1.3	1.2	1.2	1.1	0.9	0.8	0.9	1.3	1.3	0.9
Some college/Assoc deg	4139		1.0	0.9	1.0	0.8	0.9	1.0	0.9	0.9	1.0	0.9	1.0	1.0	0.9	1.0	0.8	0.8	0.8	0.8	1.0	1.0	0.9
Bachelor's degree	2986		0.8	0.7	0.9	0.7	0.8	0.8	0.8	0.7	0.8	0.7	0.7	0.8	0.8	0.8	0.7	0.8	0.6	0.8	0.9	0.9	0.8
Master's/Doctoral degree	2149		0.8	0.7	0.8	0.6	0.8	0.8	0.8	0.7	0.8	0.7	0.6	0.9	0.7	0.8	0.6	0.7	0.6	0.6	0.8	0.8	0.7
			T	S	PF	P	A	D	E	T	S	PF	P	A	D	E	T	S	PF	P	A	D	E
Metastatic status		.192
Not metastatic	4942		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
Metastatic	6708		1.6	1.1	1.2	1.6	1.2	1.3	1.3	1.5	1.1	1.7	1.9	1.2	1.3	1.6	1.4	1.1	1.2	1.4	1.1	1.1	1.1
Uncertain	2940		1.2	1.1	1.2	1.3	1.1	1.0	1.0	1.3	1.3	1.4	1.3	1.2	1.2	1.4	1.4	1.0	1.1	1.4	1.2	1.2	1.0
Payor		.150
Government	8894		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
Non-government	5663		0.9	0.9	0.9	0.7	0.8	0.9	0.9	0.8	0.9	0.8	0.7	0.8	0.8	0.8	1.1	1.1	0.9	1.0	1.2	1.0	1.0
Portal use in past 365 days		.113
No portal use	3618		-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
≥1 portal use	10,972		0.8	0.9	0.8	0.9	0.9	0.9	0.8	1.0	1.0	1.0	0.9	0.9	0.9	1.0	1.0	0.9	0.9	0.9	0.9	0.9	0.9
Cancer sitec
Breast	3503	0.108	1.0	1.4	1.5	0.8	3.0	1.3	0.6	1.1	0.4	0.5	0.7	1.1	1.8	0.5	0.4	0.5	0.6	0.5	1.2	0.6	0.2
Endocrine	909	0.072	1.1	1.4	1.2	0.6	2.9	1.4	0.6	1.5	0.5	0.7	0.9	1.7	2.3	0.8	0.5	0.5	0.6	0.6	1.4	0.8	0.2
GI	3804	0.074	1.3	1.4	2.0	0.8	3.1	1.5	0.8	1.2	0.5	0.6	0.6	1.3	2.2	0.7	0.5	0.5	0.7	0.7	1.3	0.8	0.3
GU	1998	0.071	1.1	1.3	1.8	0.8	2.3	1.3	0.7	1.5	0.6	0.8	0.8	1.3	2.3	0.8	0.6	0.5	0.7	0.7	1.4	0.8	0.3
Gyn	1272	0.081	1.0	1.1	1.5	0.6	2.6	1.1	0.6	1.1	0.4	0.5	0.6	1.3	1.9	0.6	0.6	0.5	0.9	0.6	1.3	0.7	0.3
Head and neck	872	0.064	1.2	1.4	1.3	0.8	2.2	1.3	0.6	1.0	0.4	0.4	0.8	1.3	2.2	0.5	0.6	0.7	0.6	0.9	1.5	0.8	0.3
Heme	1428	0.067	1.1	1.4	1.6	0.9	3.1	1.6	0.7	1.3	0.5	0.8	0.9	1.2	2.0	0.8	0.5	0.5	0.7	0.7	1.1	0.7	0.3
Lung	2205	0.153	1.1	1.3	2.0	0.8	3.5	1.6	0.8	1.6	0.5	0.9	0.9	1.7	2.8	0.9	0.7	0.7	0.8	0.8	1.7	1.0	0.4
Melanoma	678	0.067	0.9	1.1	1.1	0.6	2.6	1.3	0.7	1.0	0.4	0.4	0.5	1.1	2.1	0.5	0.4	0.4	0.4	0.4	1.1	0.6	0.2
Nervous system	1256	0.105	1.7	1.1	2.2	0.6	4.3	2.2	0.9	1.4	0.4	0.9	0.7	1.4	2.7	0.9	0.4	0.5	0.6	0.5	1.2	0.7	0.3
Sarcoma	913	0.050	1.1	1.3	1.9	0.8	3.1	1.5	0.5	1.3	0.4	0.7	0.9	1.4	2.1	0.7	0.4	0.6	0.7	0.6	1.0	0.7	0.3
Skin	1789	0.095	0.9	1.1	1.6	0.8	2.2	1.2	0.6	1.0	0.4	0.6	0.7	1.1	1.7	0.6	0.5	0.5	0.7	0.7	1.4	0.8	0.3
Multiple	6075	0.056	0.9	0.8	0.7	1.2	0.4	0.7	1.5	0.8	2.1	1.6	1.4	0.7	0.5	1.4	1.9	1.9	1.7	1.6	0.8	1.3	3.8

Abbreviations: T = SPPADE composite score. S = sleep. PF = physical function. P = pain. A = anxiety. D = depression. E = low energy/fatigue

^aThe reference group for the 3 symptom trajectories shown in the table is the Low Stable symptom trajectory. For each variable, the relative risks are compared to the first subgroup listed for that variable, except the reference group for each cancer type are individuals without that type of cancer. Each relative risk for a variable is adjusted for all other variables in this table

^bSMD = standardized mean difference where an SMD ≥ 0.20 for a particular variable indicates it may be a factor associated with symptom trajectories. The SMD shown in table is the average SMD across the 7 symptom scores

^cNone of the specific cancer sites had an SMD suggesting trajectory imbalance. The SMD range across sites was 0.06 to 0.15 with the mean being 0.07

Bolded values indicate either (1) those variables for which SMD is greater than or quite close to 0.20 (thus suggesting a potentially greater influence on symptom trajectory) or (2) those symptoms for which the 95% confidence interval of the relative risk does not include 1.0 (thus indicating a specific symptom trajectory for that particular symptom differs from a stable low trajectory at p < 0.05

In the multivariable multinomial modeling, the reference trajectory was low stable and for each variable, there was a reference category. Thus, each relative risk (RR) shows the risk of a specific variable category (compared to the reference subgroup) for a specific symptom trajectory (compared to the low stable trajectory). RRs significant at p < 0.05 are bolded in Table 4 (meaning that the 95% confidence interval does not include 1). For example, the bolded RR of 1.5 for worsening of physical function among patients aged 40 to 64 means that a worsening trajectory is 1.5 more likely than a low stable trajectory for physical function in this middle-aged group than in patients younger than 40. Conversely, the bolded RR of 0.6 means that individuals in the 40–64 age group are 0.6 times as likely (or 40% less likely) to follow a high stable anxiety trajectory compared to the younger group, relative to the comparator of a low stable trajectory. With the reference trajectory always being low stable, the other 3 trajectories represent patients who at some point in their course have elevated symptom severity which either persists (high stable), progresses (worsening), or declines (improving). Therefore, an alternative binary conceptualization would be an Elevated (3 trajectories) vs. Non-elevated (1 trajectory) grouping. Interpreting the RRs in Table 4 in this way means a significant (bolded) RR > 1 indicates that subgroup has a higher likelihood of an elevated symptom trajectory compared to the reference subgroup, whereas a bolded RR < 1 indicates a lower likelihood of an elevated trajectory.

The RRs for a few characteristics for which the SMD did not suggest imbalance are worth noting. First, cancer site was generally not associated with membership in a symptom trajectory, with the exception of 2 symptoms: fatigue improvement was less likely than a low stable level of fatigue across all cancer sites (except those with multiple cancers), and a stable high level of poor sleep was less likely than a low stable level across several cancer sites. Second, the RR are significantly greater than 1.0 for multiple symptoms in patients who are African-American or Hispanic; however, the sample size of these two subgroups is quite small (n = 132 each), necessitating a cautious interpretation of these findings.

Discussion

This trajectory analysis of SPPADE symptoms in patients with cancer revealed several important findings. First, using rigorous methods we consistently identified 4 trajectory patterns over time across 6 commonly occurring symptoms, with the mean sample proportions across all symptom scores being 65.5% for stable low, 15.5% for stable high, 10.7% for improving, and 8.4% for worsening. Second, the symptom trajectory patterns were remarkably similar across SPPADE symptoms, suggesting common underlying dynamics of symptom stability and change. Third, being employed, having higher educational attainment, being married or partnered, being 65 years or older, and not having metastatic cancer were associated with a significantly greater likelihood of being in the low-stable trajectory group. Cancer site was generally not associated with symptom trajectory except fatigue was less likely to show improvement across all cancer types. Several study strengths are the large sample size (n = 14,590 with at least 3 symptom assessments in a 12-month period), inclusion of all cancer sites and stages, and the innovation of using non-panel EHR data.

Previous trajectory analyses have most often assessed only 1 or 2 symptoms (e.g., depression, fatigue, sleep) in a single cancer site or stage. Few studies have examined multiple symptoms across multiple cancer sites. For example, two Canadian studies used the 9-item Edmonton Symptom Assessment Scale which includes 5 of the 6 SPPADE symptoms. However, one of these studies focused exclusively on metastatic breast cancer [5], and the second study focused solely on symptoms during the last 6 months of life [6]. Another study examined the 12-month course of 22 somatic symptoms in 405 patients with different cancer sites and stages [7]. However, the study did not use latent class growth analysis, and included only patients who had baseline depression and/or pain. Chae et al. used EHR data from 204 patients on the 8 symptoms most commonly documented and found that neural network and random forest models trained on previous symptom experiences predicted the subsequent trajectory better than the last clinical observation [8]. However, specific trajectories were not reported and the sample only included patients following chemotherapy. Besides Chae et al., our study is among the few to use real-world EHR data for trajectory analysis.

We integrated all 4 trajectory classes and all 6 SPPADE symptoms along with multiple baseline variables in our analysis. Presenting both the composite score trajectory as well as individual symptom trajectories provides a comprehensive view of symptom patterns in real-world practice. At the same time, complementary research might include more detailed investigation of non-SPPADE symptoms, selected trajectory classes, and particular sites and/or stages of cancer along the cancer control continuum. Research foci could include (1) predictors of improvement in individuals with high symptom severity or predictors of worsening in those with low levels of symptoms; (2) treatment-related symptom trajectories during or following chemotherapy, radiation therapy or surgery; (3) symptom trajectories in patients with advanced cancer; (4) evaluating additional predictors such as laboratory values and medication data; and (5) incorporating more granular markers of cancer stage besides metastasis. Methods from E2C2 may be adapted in future real-world studies presuming oncology practices continue to routinely collect longitudinal symptom data in the health record.

Single-item numeric rating scales are public domain and have proven to be effective tools to identify and monitor symptoms [19–22, 30]. These features, coupled with the ability to capture symptom scores using self-report, automation, and flexible modes of patient-reported outcome administration (in-person or virtually) make symptom monitoring a promising low-cost tool for clinical surveillance. Additionally, collaborative care telehealth-delivered symptom management has proven effective in both the E2C2 trial and previous studies [4, 31, 32]. Thus, symptom screening, monitoring and management can be conducted remotely, thereby enabling longitudinal data collection without requiring in-person visits.

Individual SPPADE symptom trajectories could be used to guide a more individualized approach to symptom surveillance and management. For example, patients who are consistently in the low stable trajectory may not require intervention or frequent symptom surveillance, allowing resources to be directed to those in the stable high or worsening trajectory groups. Those with less severe symptoms may be initially offered self-management resources as used as step 1 in E2C2² or other web-based symptom management programs [33]. Additionally, approaches using EHR data to identify symptoms and predict trajectory are emerging [8, 34]. Given the prominence of education, marital status and employment (possibly indicating underinsurance or lower social support or financial resources) as predictors, enhanced care management for those with lower social support or greater financial distress should be studied.

Our study has several limitations. The beginning of the trajectory was defined by first symptom screener completion in E2C2 rather than at the more conventional surveillance start times of initial diagnosis, treatment onset, or recurrence. Second, symptom data were gathered at varying intervals (depending on visit frequency and patient willingness to complete optional surveys) rather than uniform timepoints. However, importantly, latent class growth models are robust to varying intervals of measurement. Third, the multinomial analyses identify characteristics associated with trajectory-class membership but do not constitute or validate a prospective prediction model. While we therefore cannot make strong causal inferences, the temporal relationship between baseline factors and subsequent trajectories makes it much more likely that the former influence the latter rather than vice versa. Fourth, because patient assessments were derived from control, intervention, or both periods, the intervention effect on trajectory is indeterminate. Fifth, because E2C2 survey respondents were predominantly White and well-educated, the generalizability of our findings to more diverse populations requires additional study. The respondents also included a high proportion of patients with metastases. Sixth, all study patients had cancer, and consequently we were unable to estimate the effect of having any cancer to having no cancer. This may have obscured an association between cancer site and trajectory class membership. Seventh, although full maximum likelihood estimation accommodates missing data, attrition and uneven response patterns may have influenced trajectory estimation. Finally, our dataset did not include some objective clinical variables (e.g., weight, nutrition, physiological indicators) which might affect symptom burden as well as symptom trajectory.

In conclusion, two-thirds of patients with SPPADE symptoms have stable low severity symptoms; this group might warrant further research to determine whether particular subgroups of individuals with a low stable trajectory warrant monitoring and self-management programs. Adverse prognostic factors associated with higher severity symptoms at some time point include metastatic disease, lower educational attainment, being disabled or unemployed, and being not married or partnered. Oncology practices that routinely use patient-reported symptom screeners can use EHR data to tailor evidence-base symptom management strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Andrea Cheville, Jeph Herrin, Veronica Grzegorczyk, and Minji Lee. The first draft of the manuscript was written by Kurt Kroenke and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

E2C2 is funded as part of the Improving the Management of symPtoms during and following Cancer Treatment (IMPACT) Consortium, a National Cancer Institute Cancer Moonshot™ Research Initiative under the authorization of the 2016 United States 21 st Century Cures Act. This research was supported by the National Cancer Institute of the NIH, UM1CA233033 (PI Cheville, Mayo Clinic, Rochester, MN). The findings expressed in this manuscript do not necessarily reflect the opinion of the NIH.

Data availability

With publication, deidentified participant data, as well as the data dictionary and study protocol will be made available upon request to cheville.andrea@mayo.edu.

Declarations

Conflict of interest

The authors declare no competing interests.