Laboratory tests as short-term correlates of stroke

Several stroke risk scores have been developed to identify those with the highest risk of stroke [1‐4]. These stroke risk scores are mostly based on information one could collect when taking a patient’s medical history. Identification of those most at risk for developing stroke would allow focused education on both reducing risk factors and recognizing signs of a stroke. Early recognition and treatment of stroke can substantially reduce both the direct and indirect costs of a stroke [5].

Electronic health records (EHRs) are rapidly being adopted by medical providers, and are now used by the majority of office-based physicians and hospitals in the US [6, 7]. These EHRs do not always provide easily computable information regarding a patient’s medical history, but do an excellent job of providing discrete data from laboratory tests, imaging studies, and pharmacy records. Thus an opportunity exists to develop a stroke risk score that one could compute from discrete data contained in EHRs.

The best known stroke risk score is the Framingham Stroke Risk Profile, developed as part of the Framingham Heart Study [1, 2]. This risk score was based on an analysis of 472 stroke events. The score computes the probability of developing a stroke within the next 10 years based on age, sex, systolic blood pressure, along with categorical factors disclosed by the patient when taking a medical history such as whether the patient smokes cigarettes or has been diagnosed with diabetes or atrial fibrillation.

Two other stroke risk scores have been published. First, a short term (3 year) measure of stroke risk was developed based on 188 strokes observed in data collected by the Cardiovascular Health Study [8, 9]. Second, a long term (two decades) stroke risk score was developed based on 282 strokes observed in a population of municipal employees in Israel as part of the Israel Ischemic Heart Disease Project [4].

Several different health outcomes have been predicted from electronic health records [10]. Examples include: the identification of drug-drug interactions [11, 12], computation of the genetic risk for diabetes [13], identification of diabetes medication that significantly increased risk of myocardial infarction [14] and prediction of the patients’ future risk of receiving a diagnosis of domestic abuse [15].

The purpose of this study is to identify laboratory tests that effectively correlate with the occurrence of stroke. This study is based on 26,954 observations of stroke in a one-year period from the electronic health records of a large managed care organization.

We tested which laboratory tests were correlated with a future diagnosis of stroke using Cox Regression, controlling for age and sex differences. We used a database of electronic health records from Kaiser Permanente containing one year of patient records containing records for 2.4 million patients, 26,964 of whom received a diagnosis of stroke during the one year period. We compared laboratory test results that the 26,964 patients received several months before their diagnosis of stroke with the laboratory test results that the patients who never received a stroke diagnosis. From this comparison, we identified laboratory tests that had significantly different results in the two populations: those who would have a stroke within a few months and those who would not.

Kaiser Permanente (KP) is a non-profit health plan with 9.6 million members, and the largest region is Southern California with over 3.8 million members. KP offers comprehensive health care including outpatient and inpatient care, laboratory services, and pharmacies. Kaiser Permanente implemented electronic health records before 2009 [16, 17].

This is a retrospective data-only study. All data for this study were selected from members of Kaiser Permanente’s Southern California Health Plan who had continuous coverage over a recent one-year time span inclusive of 2013. Subjects were not systematically tested, but only received specific tests indicated by their medical condition. We limited the subject pool to only those patients that had received any form of care from Kaiser Permanente during 2013, including laboratory-only visits and both inpatient and outpatient encounters. Furthermore, we included only one laboratory result per type of test per month. We limited the data to one result per month to prevent over counting individuals who may be sampled more frequently, which would bias our results. Only discrete data elements of demographics, diagnoses, laboratory tests, and EKGs were extracted and analyzed, from which all Protected Health Information (PHIs) were removed during data extraction.

Overall, the study included 2,412,213 individuals. Of these 51 % (1,239,559) were female. The birth-year distribution of this population is shown in Fig. 1 and the frequency of self-reported race/ethnicity categories are shown in Additional file 1: Table S1.

We identified 38 tests that effectively correlate with stroke. These 38 tests included three from electrocardiogram measurements, two from urine tests, and 33 from blood, serum and/or plasma. Many of the 33 blood tests are part of standard panels, including 12 that are typically included in (or calculated from) a comprehensive metabolic panel and 14 from a complete blood count with differential and 5 that are included in a lipid panel. The other two blood tests are used to diagnose heart failure (natriuretic peptide B, also known as BNP) and to monitor tumor progression (carcinoembryonic antigen, also known as CEA).

These 38 tests that proved to be correlated with stroke both reproduced previously known predictors of stroke and introduce several novel associations. The previously known correlates include kidney dysfunction and atrial fibrillation [25, 26].

We identified several measures of kidney function as significant correlate of stroke. Glomerular filtration rate is a standard measurement of renal function [27]. It is a calculated value based upon serum levels of creatinine, urea nitrogen, and albumin along with demographic factors: age, sex, and race [28]. Creatinine has been previously recognized as a stroke predictor [29, 30].

The stroke prediction model [3] based upon the Cardiovascular Health Study [9] includes only one laboratory value, serum creatinine levels. The authors divided the population into two groups, split with serum creatinine levels above or below 1.25 mg/dL. In contrast, we found serum creatinine levels to be a strong correlate with continuous range from less than 0.58 mg/dL to greater than 1.74 mg/dL as shown in Figs. 3, 4 and 5.

The BNP test is used to diagnose congestive heart failure [31]. High levels of BNP (greater than 480 pg/ml) suggest congestive heart failure, while low levels (less than 100 pg/ml) rule it out [32]. BNP has been shown to predict mortality [33, 34]. We found BNP to be a particularly good correlate of stroke even at low and intermediate levels. We saw significant differences in the prognosis for stroke even between those who had BNP levels less than 13 pg/ml and those who tested less than 34 pg/ml.

Three parameters measured by electrocardiograms also appear in Table 1, including the Q-T interval, the QRS duration, and the T wave axis. Increases in both Q-T interval and QRS duration were positively correlated with stroke incidence. T wave axis was also more positively correlated as the axis shifted rightward (i.e., increased).

Of the 38 tests that were correlated with stroke, many fell into distinct groups of similar function. Four laboratory tests associated with diabetes all correlated positively with stroke, which supports previous research on the subject [35]. Similarly, laboratory results that might reflect impaired kidney function, such as creatinine, urea nitrogen, and potassium, are also positively correlated with stroke chance. This also agrees with previous work showing that patients with Chronic Kidney Disease are associated with a higher risk of stroke [36].

Interestingly, higher levels for non-HDL cholesterol are significant risk factors for cardiac disease [37], but not necessarily for stroke. The MRFIT study of 350,977 men over six years found that lower levels of total cholesterol lead to a higher risk of death by intracranial hemorrhage, while higher levels did lead to an increased risk of non-hemorrhagic stroke [38]. Men with serum cholesterol levels less than 160 mg/dL had three times the risk of intracranial hemorrhage compared to those with higher cholesterol levels. The more recent Physicians’ Health Study, 22,071 men followed for 11 years, found no significant correlation between total cholesterol, HDL cholesterol, or triglycerides and ischemic stroke [39].

Our findings showed that higher levels of total cholesterol, HDL, LDL, and non-HDL cholesterol were correlated with significantly decreased incidence of stroke. Although other research has shown that high HDL is correlated with decreased risk of stroke, the research on total cholesterol and LDL has been more mixed. Our results for these tests are highly significant, and this finding is interesting because many of the patients are likely treated with statins, which are known to reduce stroke risk while lowering LDL and total cholesterol [40]. It is possible that stroke risk depends on long-term cholesterol levels, and the use of statins suddenly reduces cholesterol levels, thus distorting this data. However, since we did not include medications in our assessment, we are unable to determine what relationship treatment might have on interpreting these findings.

The significance of the EKG correlations in our results in unclear because the clinical assessments of the EKGs were not available in the discrete data analyzed. It is possible that increases in the Q-T interval and QRS duration being correlated with increase in stroke was due to underlying cardiac disease and secondary cardiac manifestations of that disease like bundle branch blocks or the use of antiarrhythmic medication. The positive correlation of rightward axis deviation is possibly due to similar causes, or possibly associated with COPD and right-sided heart failure or right ventricular hypertrophy. Further assessment of this relationship would help clarify the significance of this correlation.

The results for iron-related labs are inconsistent. High TIBC is negatively correlated with stroke, but high iron is also negatively correlated, although at a much lower significance level. Since high TIBC is a better correlate, it could imply that iron deficiency is correlated with increased risk of stroke, although we would also expect low iron in that situation. The literature on the association of iron with stroke appears to be mixed, with a recent study [41] on a specific population of patients showing positive correlation, and other studies showing negative [42] or mixed [43] correlations. Further analysis of iron’s association with stroke may be helpful.

Albumin was the strongest negative correlate for stroke in our results. This observation is consistent with numerous studies that showed a beneficial effect of high albumin on risk of stroke [44‐46].

We identified common laboratory blood tests that correlate with stroke. Some were positively associated (neutrophil count and percent, CD3 + CD8+ T8 suppressor cells, monocytes, eosinophils, and CD3 cells), while others were negatively correlated (hematocrit, lymphocytes). Given the numerous clinical reasons for these blood tests to be both high and low across inpatient and outpatient encounters, it is difficult to determine any likely justification for the correlation from these readings.

This analysis has a number of limitations. The discrete data used is only from a one-year time period, and may benefit from expanding to additional years. We selected for review only patients with KP health insurance that was continuous, thereby limiting the assessment to only insured patients. This selection likely underrepresents for people that were uninsured or partially insured for any reason. We did not include claims data to identify KP members that were diagnosed with strokes outside of the KP network of care or outside of this time period, and therefore may not include all KP members that had strokes during that timeframe or before the beginning of the timeframe. Our method of taking the median laboratory value when multiple results exist during a month may artificially normalize abnormal readings, doesn’t fully take into account the potential impact of acute care events, and even then isn’t necessarily independent. Other methods exist for considering time correlations that could be used in a similar assessment [47].

We did not include medication data in our analysis, which could have a significant impact on some of our findings. In particular, the significance of the cholesterol tests on stroke is likely complicated by the high rate of LDL control using cholesterol lowering medications, like statins, at KP, and their ability to reduce the risk of strokes [40, 48]. Additional analysis that factored in medication use would likely raise additional findings for consideration.

We included both hemorrhagic and non-hemorrhagic strokes in our analysis, and did not differentiate between the two. This allowed us to compare our results to other landmark studies [1, 2]. Since the risk factors for these two conditions are different, and the number of non-hemorrhagic strokes is significantly higher than hemorrhagic strokes, our analysis may not accurately reflect the correlation between these lab tests and hemorrhagic strokes. Further assessment of each sub-group would be an important area for further analysis. Over 95 % of the strokes analyzed were non-hemorrhagic.

We used ICD-9 diagnoses documented in the medical record to identify patients with a stroke. Since this included problem list diagnoses, it is possible that patients were included that had a history of stroke and not a current stroke event if that diagnosis was added to the problem list during our timeframe. Also, a diagnosis of stroke is only appropriate as an acute diagnosis, but we did not limit our selection to only acute encounters, and therefore could have over-selected patients into our cohort who had previously had strokes. We could have limited our selection of patients to only first-time stroke patients using more historic data, and not including patients with diagnoses representing side effects of strokes, but did not do so.

In conclusion, we identified several dozen independent laboratory tests that are strong correlates of stroke. These laboratory tests could be combined to provide a short-term (one year) measure of the probability that a patient will be diagnosed with a stroke.

BNP, B-type natriuretic peptide; CD3, cluster of differentiation; CEA, carcinoembryonic antigen; COPD, chronic obstructive pulmonary disease; EKG, electrocardiogram; HDL, high density lipoproteins; ICD, International Classification of Disease; KP, Kaiser Permanente; LDL, low density lipoproteins; LOINC, logical observation identifiers names and codes; MRFIT, multiple risk factor intervention trial; PHI, personal health identifiers; TIBC, total iron binding capacity