Evaluation of the Use of Electronic Health Data to Classify Four-Year Mortality Risk for Older Adults Undergoing Screening Colonoscopies

Abstract

Current cancer screening recommendations often apply coarse age cutoffs for screening requirements without regard to predicted life expectancy. Using these cutoffs, healthier older patients may be under-screened, and sicker younger patients may be screened too often. Mortality risk classification using EHR data could be used to tailor screening reminders to physicians in ways that better align screening recommendations with patients who are more likely to live long enough to benefit from early detection. We have evaluated the performance of an existing prognostic index for 4-year mortality using data readily available in the electronic health record (EHR), and investigated the effect of the index in retrospective cohorts of adults age 65 and older undergoing screening colonoscopy. Risk scores in this adaptation of a four-year prognostic index were found to be associated with actual death rates and consistent with mortality rates from a national sample. Our results demonstrate that data extracted from electronic health records can be used to classify mortality risk. With improvements, including extension to a 5-year mortality model with inclusion of additional variables and extension of variable definitions, informatics methods to implement mortality models may prove to be clinically useful in tailoring screening guidelines.

Introduction

The goal of cancer screening is to detect the presence of neoplastic disease at a stage when intervention is expected to prolong life. It is not effective to screen a patient who is very likely to die of causes other than the disease being screened. Inappropriate screening presents problems from a resource utilization standpoint – screening an individual who is not likely to benefit delays the screening of people who may receive more benefit. Also importantly, the risk/benefit ratio in people who are not likely to benefit from screening may be too high, and direct negative consequences of the procedure may do more harm than the potential benefits of screening. Most guidelines recommend initiation of colon cancer screening in low risk individuals starting at age 50, with repeat screening every 10 years or sooner if the initial colonoscopy has adenomatous polyps. However, there is divergence of opinion on if and when screening should stop. The US Preventive Services Task Force (USPSTF) and ACS-MSTF American Cancer Society–U.S. Multi-Society Task Force (ACS-MSTF) both released updated guidelines on colon cancer screening in 2008. The USPSTF guidelines recommend against routine colorectal cancer screening from age 75 to 85 years, except in special circumstances, and recommend against routine screening after age 85 years under any circumstances¹. The ACS-MSTF guidelines do not provide any age related recommendations against screening². The conflict between the guidelines and the topic of when to stop screening has been raised in a recent editorial in the Annals that called for consistent guidelines providing specific starting and stopping ages³. However life expectancy rather than chronological age has been found to be a deciding factor in receiving benefits from early detection of cancer, with life expectancy being a function of age plus comorbidities, severity of disease, and functional status.^{4,5,6,7,8,9,10} Randomized trials of colon cancer screening found that it took 5 years before screened patients had a clearly lower mortality rate from colorectal cancer than control participants. Although age alone may be an important predictor in a model of life expectancy, other clinical conditions may also have relevance in assessing life expectancy and therefore the appropriateness of colon cancer screening. In a recent review of the available evidence on the effects of comorbidity on cancer screening in elderly individuals, Terret et al11 have proposed that decisions on cancer screening in older adults should be based on a comprehensive geriatric assessment to help avoid under-screening in healthy people and over-screening in frail individuals. The performance of this type of assessment frequently involves an interdisciplinary team of providers, and is not available to assist primary care providers in making point-of-care screening recommendations. As an alternative, secondary use of data from Electronic Health Records (EHRs) have the potential to provide real-time tailoring of cancer screening guidelines. An electronic health record-based prognostic index could benefit patient care by providing point-of-care life-expectancy estimates that are individualized the each patient’s health status, and can be used to tailor screening reminders to physicians and improve adherence with process-based quality of care measures.

In this paper, we report on the evaluation of the performance of an existing prognostic index for 4-year mortality using data readily available in the electronic health record. We also describe the potential effect of applying the index in two retrospective cohorts of adults age 65 and older undergoing screening colonoscopy to see if important colonoscopy findings would have been missed if the mortality index resulted in a recommendation against screening.. The primary aims of our investigation are described in the following sections:

Evaluate an EHR Based Adaptation of a Prognostic Index for Four-Year Mortality in Older Adults

The purpose of this component of the study was to evaluate the performance of an existing prognostic index for 4-year mortality¹² in a retrospective cohort of older adult primary care patients, appropriately adapted to leverage data readily available in the EHR. A retrospective cohort of older adult primary care patients was scored for risk of 4-year mortality according to the model of Lee, et al¹² which had been developed on a national sample of patient self-report data from the 1998 Health and Retirement Study. The clinical and demographic predictor variables and related mortality risk points are shown in Table 1. Mortality associated with different point totals for the national sample is shown in Table 2.

Table 1.

1998 Health and Retirement Study Mortality Risk Index¹²

Variable	Points	Variable	Points
Age 65–69	2	Smoker	2
Age 70–74	3	Diabetes	1
Age 75–79	4	Cancer	2
Age 80–84	5	COPD	2
Age 85+	7	Heart Failure	2
Male	2	Dementia	2
BMI <25	1	Weakness	5

Table 2.

1998 Health and Retirement Study Mortality by Risk Group¹²

Risk Group (points)	%Mortality
1 (0–5)	3 – 4%
2 (6–9)	15%
3 (10–13)	40 – 42%
4 (14+)	64 – 67%

Evaluate Four-Year Mortality Risk, Four Year Survival, and Colonoscopy outcomes during 2003–2007 for adults age 65 and older who were residents of Pennsylvania

The purpose of this component of the study was to assess the potential impact of basing recommendations for colonoscopy screening on mortality risk through use of a EHR-based program that would provide estimated risk of death within 4 years to primary care providers. Patients undergoing elective colonoscopies are stratified by age groups and mortality risk score. The subsequent pathology reports were stratified into positive and negative findings. Given the more recent time interval under observation for this part of the study, we could not assess the accuracy of the calculated mortality rate. Instead, we sought to assess the use and outcomes of colonoscopy among people having different mortality risks and determined the extent to which important colonoscopy findings may be missed in people who would have been excluded from a colonoscopy under our new reminder system that bases its recommendation on mortality risk.

Evaluate Four-Year Mortality Risk and Colonoscopy outcomes during 2009 for adults age 65 and older

The purpose of this component of the study was to conduct further investigation of the potential impact of mortality risk models in clinical decision making for colonoscopy screening. The analyses is developed on clinical and colonoscopy data collected in 2009 with the objective of developing risk scores that are based on more complete clinical data and to examine current clinical practices.

Methods

For both data availability and temporal reasons there was not a single data source currently available in the University of Pennsylvania Health System (UPHS) sufficient to address the multiple study requirements of adequate clinical data, colonoscopy reports, pathology findings, and mortality data for both four-year follow-up and recent time periods. At the time the study was done in 2010, a full assessment of 4 year mortality required source data from prior to 2005 to ensure a minimum of 4 years of follow-up. However, this earlier time period coincides historically with the start-up of the large scale conversion to the EHR at UPHS. Although many patients were receiving colonoscopies prior to 2005, the overlap between patient populations represented in the EHR and the endoscopy system in this time period was limited. To address the potential impact of the utilization of the mortality model based on current practice, recent data from the endoscopy system would have been ideal. However the data access capabilities of the more recent version of the endoscopy system paradoxically are more limited than the full reports that were available in the older system. For these reasons, a multi-pronged approach to sampling and conducting the analyses was taken as a means of working around the limitations of individual data sources.

Data Sources

The systems used to compile data for this study included the UPHS EPIC EHR system, billing systems, laboratory system, endoscopy system, surgical pathology system, and Pennsylvania Department of Health mortality files. Demographic and clinical data was primarily drawn from the EHR, and additional data from billing and lab systems was incorporated to overcome limitations stemming from the early years of EHR implementation. The endoscopy and surgical pathology systems were used to determine colonoscopy findings, and state mortality files used for classification of four-year survival.

Patient Populations

The cohort for evaluation of the four-year mortality risk model was drawn from the EHR and included all adults age 65 and over at time of last encounter with a UPHS primary care practice (regardless of colonoscopy status) between 01/01/2003 and 12/31/2004, who were also residents of Pennsylvania. This analysis is limited to Pennsylvania residents as survival outcomes are determined through Pennsylvania Department of Health death data. The clinical and demographics characteristics of the cohort, as well as the risk point scores assigned to each category based on the Health and Retirement Study Mortality Risk Index¹² study are shown in Table 3.

Table 3.

EHR Risk Factor Index

Risk Factor	Risk Factor Definitions	N (%)	Points
Age	65 – 69	2240 (27.7)	2
	70 – 74	1984 (24.5)	3
	75 – 79	1692 (20.9)	4
	80 – 84	1218 (15.1)	5
	>=85	956 (11.8)	7
Male Sex	Demographics	3049 (37.7)	2
Diabetes	250.[1–9]* Diabetes with complications or manifestations	2563 (31.7)	1
	HbA1C >10
	CRE >2
Cancer	140 - 209.30 Malignant Neoplasms	1062 (13.1)	2
	(exclude “173*” – Basal Cell Cancer)
Lung Disease	490–496 COPD And Allied Conditions	888 (11.0)	2
Heart Disease	398.91 Rheumatic heart failure (congestive)	1162 (14.4)	2
	402.01 Hypertensive heart disease with heart failure
	404.*1 Hypertensive heart and chronic kidney disease with heart failure
	404.*3 Hypertensive heart and chronic kidney disease with heart failure
	428.* Heart failure
	430–438 Cerebrovascular Disease
BMI <25	Clinical height & weight data	2095 (25.9)*	1
	783.2 Abnormal loss of weight and underweight
	783.22 Underweight
Current Smoker	Social history of current smoking	749 (9.3)*	2
Cognitive	290.* Dementias	495 (6.1)	4
	294.1* Dementia in conditions classified elsewhere
	331.* Other cerebral degenerations
Motor	719.7 Difficulty in walking	792 (9.8)	3
	728.87 Muscle weakness (generalized)
	780.7* Malaise and fatigue
		Total N 8090	Max Points 26

^*BMI missing for 13.7% of participants, Smoking History missing for 15.2% of participants

To create the second study cohort that looked at the utilization and findings of colonoscopy as a function of the 4 year risk score, we utilized the endoscopy database to identify patients age 65 or older undergoing elective colonoscopies between 01/01/2003 and 12/31/2004. This analysis is limited to Pennsylvania residents since survival outcomes are determined through Pennsylvania Department of Health death data.

The cohort for third assessment of current colonoscopy risk profiles was drawn from endoscopy appointment scheduling data. It is not known in this cohort if the colonoscopies were for screening or part of a work up for GI symptoms, but patients with a known prior history of colon cancer or inflammatory bowel disease were excluded from the sample.

Measures

The EHR adaptation of the four-year mortality risk model used directly-equivalent variables from the electronic health record when available (age, gender, smoking status, BMI), or substituted related historical ICD-9 codes and laboratory values consistent with the diagnostic components. Risk factors were coded by examining data for one year prior to last study encounter within the observation period. The date of last encounter within the health system was determined by linkage of EHR records to all health system billing encounters via a master patient index. Participants with fewer than four years follow-up within the health system, suggesting they were either dead or lost to follow-up, were then submitted to the Pennsylvania State Health Department for matching to mortality files. The model variables, definitions, and related risk score points are shown in Table 3.

Results

Evaluation of EHR Based Adaptation of a Prognostic Index for Four-Year Mortality in Older Adults

Risk scores were calculated and four-year mortality determined for a sample of 8090 patients. Mortality increases with age from 10% in the lowest age group to a high of 53.2% in those aged eighty-five and above (Table 4). The local age-related mortality rates are consistent with mortality rates from the national sample. Increasing mortality risk score is also associated with increasing mortality, ranging from 8% mortality for those receiving five or fewer risk points to a high of 69.9% for those receiving a risk score of 14 or greater. This compares to an overall mortality of 21.9% for the entire sample of adults aged sixty-five and older (Table 5). The increasing trend in mortality associated with increasing point totals demonstrates the face validity of the model. The mortality rates associated with our general medicine population using available EHR data are consistent with those found by the Health and Retirement Study Mortality study, as outlined in Table 2. When mortality is stratified by risk score within age groups, there is a continued trend of increasing mortality within strata. For example the mortality for those aged 75–78 ranges from 14.2% for those in the lowest mortality risk group to 53.3% for those in the highest risk group.

Table 4.

Four-year Mortality by Age Groups

Age Group	% Mortality	# Deaths/Total
65–69	10.0%	225/2240
70–74	13.0%	257/1984
75–79	22.6%	382/1692
80–84	32.4%	395/1218
85+	53.2%	509/956

Table 5.

Four-year Mortality by Risk Groups

Risk Group (points)	%Mortality	# Deaths/Total
1 (0–5)	8.0%	271/3385
2 (6–9)	23.9%	827/3464
3 (10–13)	51.2%	542/1058
4 (14+)	69.9%	128/183
Overall	21.9%	1768/8090

Four-Year Mortality Risk, Four Year Survival, and Colonoscopy outcomes for Pennsylvania residents, age 65+, 2003–2004

We analyzed the mortality risk scores, colonoscopy findings, and actual 4 year mortality of 1265 patients over age 65 who underwent elective colonoscopy in 2003–2004. Key findings from the cohort are that 122 (10%) of patients undergoing elective colonoscopies had risk scores >=10, with an actual four-year mortality rate of 24.8%. The observed mortality rate in the high-risk colonoscopy patients was lower than that of the equivalent primary care population in which a comparable risk group had a 59% mortality rate. Pathology results of colonoscopies were found for 518 of the 1265 patients with 18 findings of colo-rectal carcinomas. If a cutoff score of >=10 had been applied to limit colonoscopies, only 3 out of 18 positive findings would have been missed, but, consistent with their high mortality score, 2 of the 3 high risk patients survived less than 4 years despite the colonoscopy.

Four-Year Mortality Risk and Colonoscopy outcomes during 2009 for adults age 65 and older

We analyzed the mortality risk scores and colonoscopy findings of 1829 patients who underwent colonoscopy in 2009. Key findings from the cohort are that 6.6% of patients undergoing colonoscopies had risk scores >=10 (122/1829). Pathology results of colonoscopies were found for 1081 of the 1829 patients with 23 findings of colorectal carcinomas. If a cutoff score of >=10 had been applied to limit colonoscopies, 3/23 positive findings would have been missed.

Conclusion

The relationship between an EHR-derived mortality risk score, adapted from a four-year prognostic index, and actual mortality was found to be consistent with mortality rates from the national sample that applied self-reported data to a similar risk index. Application of this index at a score cutoff of 10 points would recommend against screening 3% of people between age 65 and 74, and would favor screening 81% of people age 75–84. The lower death rate among patients with colonoscopy compared with the overall primary care population over age 65 suggests the influence of pre-existing clinical judgment that favors sending patients with a higher life expectancy for the procedure. However there remain groups of high risk younger patients undergoing screening who are less likely to benefit from treatment, as well as low risk older patients who are not being screened and could benefit. Our results demonstrate that data extracted from electronic health records can be used to classify mortality risk. With improvements including extension to a 5-year mortality model, inclusion of additional variables, and extension of variable definitions, informatics methods to implement mortality models may prove to be clinically useful in personalizing screening guidelines.