Detecting adverse reaction signals to statins in a Chinese regional healthcare database, using a tree-based scan statistic method

doi:10.21203/rs.3.rs-1611181/v1

Download PDF

Article

Detecting adverse reaction signals to statins in a Chinese regional healthcare database, using a tree-based scan statistic method

https://doi.org/10.21203/rs.3.rs-1611181/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Healthcare databases offer complementary data for drug safety signal detection, addressing the limitations of spontaneous reporting. We applied the tree-based scan statistic (TreeScan) to detect adverse effect signals to statin therapy, using 2010–2016 data from the Yinzhou regional healthcare database. Patients (> 18 years) with hypertension were included, and identified as statin users according to their out-/inpatient prescription records. Adverse events (AEs) were defined per ICD‐10 codes from participants’ diagnosis records. We detected the safety signals of statins using the TreeScan method and established a set of reference signals according to these ICD‐10 codes. The signal gold standard was constructed from published meta-analyses, systematic reviews, and statin package inserts. Diagnostic test measures—sensitivity, specificity, and positive predictive value, among others—were computed. In total, 224,187 patients were enrolled and classified as either statin users (85,758) or nonusers (138,429). A reference set of 126 ICD‐10-derived AEs, including 13 positive and 113 negative signals, was constructed; TreeScan generated 30 positive signals (p < 0.05), 9 of which were known AEs. Hence, TreeScan can be applied as a drug safety surveillance, signal detection method, simultaneously evaluating several potential AEs. It can further adjust for the multiple testing inherent in evaluating overlapping groups.

The use of statins (HMG-CoA reductase inhibitors) provides strong clinical evidence for the prevention of cardiovascular diseases (CVDs) and these drugs have been widely used in the prevention and treatment of atherosclerotic CVD(1). Despite the cardiovascular benefits of statins, studies have shown that they commonly cause adverse effects, which is the main factor discouraging patients from taking and adhering to statin therapy. Clinical investigations and systematic reviews have suggested that statins exhibit several severe adverse effects, including myopathy, rhabdomyolysis, non-serious muscle symptoms, newly diagnosed diabetes mellitus, and liver toxicity(2).

Drug safety has increasingly become a matter of serious public health concern, and measures to prevent and control drug-induced damage are urgently needed(3). Currently, drug safety surveillance is mainly based on spontaneous reporting systems(4). With their increasing availability, the use of healthcare databases as complementary data sources for drug safety surveillance, has been explored to overcome the limitations of spontaneous reporting(5, 6). The tree-based scan statistic (TreeScan) is a statistical data mining tool that has been used for drug safety signal detection(7).

However, no relevant literature reports exist on post-marketing safety studies—conducted using the TreeScan method—on statin prescription in China. Moreover, the feasibility of applying this method in the Chinese electronic medical database system, remains unclear. Thus, we sought to evaluate whether TreeScan could help identify the incidence of adverse events (AEs) associated with statin therapy. Our research involved a cohort study of adult patients with hypertension, compared to non-statin users, using data from an electronic medical database in China(8), extracted for the period between January 1, 2010, and August 31, 2016. The objective of this study was to apply the TreeScan method to detect adverse effect signals related to statin therapy and evaluate the performance of TreeScan in a Chinese regional healthcare database.

Data source and study population

The data used in our study were derived from the Yinzhou healthcare database—a linked primary and secondary care database in Ningbo City, China—for the period, 2010 to 2016(8). Patients older than 18 years with a diagnosis of hypertension, were included in the study. We identified statin users based on their out-/inpatient prescription information and AEs were defined using the ICD-10 codes on participants’ out-/inpatient diagnosis records.

Yinzhou is a district in Ningbo—an eastern coastal city in China—with a population that was nearing 1.5 million, in 2015. In 2005, the local health authority rolled out a Health Information System, which has since been collecting and administrating residents’ healthcare information. By 2010, this system had covered all hospitals in Yinzhou. Moreover, it has integrated the electronic medical records (EMRs) of hospitals, public health surveillance, and disease management, among communities. The administrative healthcare database—which has been used and described in previous studies(6, 8)—stores all data, including out- and inpatient records, prescription information, and general demographic characteristics. Out- and inpatient records comprise the hospital and patient codes, dates of diagnosis, as well as diagnosis names and codes (according to the 10th revision of the International Classification of Diseases and Related Health Problems: ICD-10). Prescription information consists of prescription codes and dates, along with drug names, brands, generic names, and dosage forms and frequencies. For this study, all patient data were anonymized, and no personal details were shared with the health researchers, ensuring patient privacy protection. The study was approved by the ethical review board of Peking University Health Science Center (approval number: IRB00001052-18013). Informed consent was waived by ethical review board because of the use of anonymized, routine data. All the used methods were conducted according to relevant guidelines and regulations.

Adverse events data mining

The potential outcomes earmarked for inclusion in our hierarchical classification system (‘tree’) were developed using the ICD-10 diagnosis codes. We screened for AEs based on these codes, assigned by physicians during participants’ out-/inpatient hospital visits. All ICD-10 codes, as well as groups of related codes at three higher levels, were evaluated as potential AEs: disease chapters (e.g., I00-I99 - Diseases of the circulatory system), disease blocks (e.g., I10-I15 - Hypertensive diseases), and three position codes. Accordingly, the ICD-10 codes defined a structured tree of diagnoses, and each grouping was defined by a cut on the tree. Each cut represented a larger or smaller group of related diagnoses, according to the structured ICD-10 codes. We excluded outcomes that were unlikely to represent adverse drug events, such as the ICD-10 chapters covering Certain infectious and parasitic diseases (A00-B99), Pregnancy, childbirth, and the puerperium (P00-P99), Congenital malformations, deformations, and chromosomal abnormalities (Q00-Q99), and Injury, poisoning, and certain other consequences of external causes (S00-T98). However, we included all other ICD-10 codes. The final tree contained 958 primary, 133 secondary, and 13 tertiary diagnoses.

Statistical analysis

First, we detected the safety signals of statins using the TreeScan method. By comparing the signals to reference signals, we could compute diagnostic test measures, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, the Youden index, and the area under the curve (AUC), according to the variables defined in Table 1.

Table 1

Contingency table of TreeScan diagnostic test evaluation
Signal detection result	Standard signals		Total
	Positive events	Negative events
Positive signal	a	B	a + b
Negative signal	c	D	c + d
Total	a + c	b + d	N

Values were calculated as follows:

$$\text{S}\text{e}\text{n}\text{s}\text{i}\text{t}\text{i}\text{v}\text{i}\text{t}\text{y}=(\text{a}/\text{a}+\text{c}) \times 100\text{\%}$$

$$\text{S}\text{p}\text{e}\text{c}\text{i}\text{f}\text{i}\text{c}\text{i}\text{t}\text{y}=(\text{d}/\text{b}+\text{d}) \times 100\text{\%}$$

$$\text{P}\text{P}\text{V}=(\text{a}/\text{a}+\text{b}) \times 100\text{\%}$$

$$\text{N}\text{P}\text{V}=(\text{c}/\text{c}+\text{d}) \times 100\text{\%}$$

$$\text{C}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t} \text{r}\text{a}\text{t}\text{e}=(\text{a}+\text{d})/(\text{a}+\text{b}+\text{c}+\text{d}) \times 100\text{\%}$$

$$\text{Y}\text{o}\text{u}\text{d}\text{e}\text{n} \text{i}\text{n}\text{d}\text{e}\text{x}=\text{s}\text{e}\text{n}\text{s}\text{i}\text{t}\text{i}\text{v}\text{i}\text{t}\text{y}+\text{s}\text{p}\text{e}\text{c}\text{i}\text{f}\text{i}\text{c}\text{i}\text{t}\text{y}-1$$

Tree-based scan statistic

We used tree-based scan statistics to identify AEs with a higher incidence. These are disproportionality statistics that adjust for multiple testing and allow for the simultaneous testing of diagnosis codes at all levels of granularity, that is, all cuts on the ICD-10 tree. An unconditional Bernoulli tree-based scan statistic was used to compare two groups; it assumes that the exposure follows a Bernoulli probability distribution. The null hypothesis was that events in all cuts would have the same probability of occurrence, in both the statin and non-statin groups. The alternative hypothesis was that events with at least one cut had a higher risk of occurrence in the statin group, than those in the non-statin group.

Because the distribution of the tree-based scan statistic is unknown, we non-parametrically derived multiple testing-adjusted p-values, using Monte Carlo hypothesis testing, where permutations of the data were generated under the null hypothesis. The multiple testing-adjusted p-value was determined by ranking the test statistics from 9,999 datasets simulated under the null and observed datasets, from largest to smallest. The p-value was calculated as the rank of the observed dataset test statistic divided by 10,000 (9,999 simulated datasets + 1 observed dataset). The multiplicity-adjusted p-values were interpreted as the probability of seeing an association of the observed magnitude, or one more extreme, if the null hypothesis were true. Together with the relative risk estimates, these p-values were used as indicators to prioritize alerts for further investigation. Cuts with a p-value below 0.05 were considered statistically significant. The analysis was performed using TreeScan v1.4 software (https://www.treescan.org) and R version 3.6.2.

Standard signals

To better evaluate performance of this method, we established a set of reference signals to identify AEs, by referring to a set of ICD-10 codes. The gold standard for these signals was constructed by searching published meta-analyses, systematic reviews, and statin package inserts(9).

Adverse drug reactions (ADRs) for which a clear association was known to exist, were used as gold standard positive event pairs. Conversely, those for which a clear association did not exist, represented gold standard negative event pairs. In this study, potential adverse drug events were identified, based primarily on ICD-10 diagnostic codes. Thus, to establish accessible drug-event pairs, ICD-10 codes that could be used as AEs, needed to be identified first. A systematic review summarized a set of ICD-10 codes for AEs, which were used as the basis for establishing relevant drug-event pairs in this study(10). These drug-event pairs were then classified as either positive or negative, based on related literature (randomized controlled trials, meta-analyses, and reviews) and drug label information(1, 2).

Characteristics and categories

A total of 224,187 patients were ultimately enrolled and divided into two groups (85,758 statin users and 138,429 nonusers). The number of patients whose out-/inpatient records reflected statin prescription during the study period, was 93,529. We excluded patients with missing data such as resident health record numbers, age, sex, and body mass index (BMI), and those younger than 18 years, retaining a total of 85,758 patients in the statin group. Patients in the non-statin group were defined as those who had no record of statin prescription during the study period and were at least 18 years old. After excluding patients with missing information, 138,429 patients remained in the non-statin group, as shown in Fig. 1.

As shown in Table 2, participants in the statin group were aged between 48 and 65 years and their average age was 57.5 years, which was higher than that of patients in the non-statin group. The standardized mean difference (SMD) exceeded 0.1, which could indicate a significant difference in age between the two groups. The proportion of female patients was slightly higher in the statin group (54.4%) than the non-statin group (50.4%); however, there was no significant difference between the two groups. There was also no significant difference between the two groups in terms of their literacy level, which mainly comprised junior high and elementary school (about 70%) education. BMI was significantly higher in the statin group than the non-statin group, with 39.3% and 34.4% of patients in the statin and non-statin groups, respectively, presenting as overweight and obese (BMI ≥ 24 kg/m²). In terms of lifestyle, the majority of patients—approximately 73% in the statin group and 72% in the non-statin group—were nonsmokers, whereas roughly 77% and 75% in the statin and non-statin groups, respectively, did not drink alcohol. Moreover, approximately 58% of patients in both the statin and non-statin groups, exercised at least 1 day per week. There were no significant differences in the levels of smoking, alcohol consumption, and exercise between the two groups (SMDs were below 0.1).

Table 2

Baseline characteristics of the study population
Characteristics N (%)	Non-statin N = 138,429	Statin N = 85,758		SMD		P
Age(year)				0.11		0.02
Mean ± SD	56.1 (13.3)	57.5 (12.2)
Median (Q1, Q3)	55.0[46.0, 64.0]	57.0 [48.0, 65.0]
Sex				0.08		0.08
Female	69812 (50.4)	46625 (54.4)
Education level				0.06		0.13
Illiterate/semiliterate	18069 (13.1)	11026 (12.9)
<9 years	101723 (73.5)	64888 (75.7)
≥9 years	13162 (9.5)	7066 (8.2)
Unknown	5475 (4.0)	2778 (3.2)
BMI(kg/m²)			0.11		0.04
<24.0	90755 (65.6)	52025 (60.7)
24.0 ~ 27.9	40498 (29.3)	27752 (32.4)
≥28.0	7176 (5.2)	5981 (7.0)
Cigarette smoking			0.06		0.19
Current	24697 (17.8)	13507 (15.8)
Never	99695 (72.0)	62884 (73.3)
Pass	10864 (7.8)	7430 (8.7)
Unknown	3173 (2.3)	1937 (2.3)
Alcohol consumption			0.04		0.27
Never	104970 (75.8)	66366 (77.4)
Often	7991 (5.8)	4771 (5.6)
Occasional	22004 (15.9)	12485 (14.6)
Unknown	3464 (2.5)	2136 (2.5)
Physical activity			0.02		0.31
>4 d/wk	59482 (43.0)	36953 (43.1)
1 ~ 4 d/wk	21308 (15.4)	13849 (16.1)
<1 d/wk	53637 (38.7)	32520 (37.9)
Unknown	4002 (2.9)	2436 (2.8)

Table 3 provides information on the statin prescriptions issued to the statin population. During the study period, a single statin prescription was recorded for approximately 28,000 patients, accounting for 32.9% of the total statin population. Issuance of 2–10 prescriptions was recorded for 44.3% of patients and more than 10 prescriptions, for 22.9% of patients. Most patients (71.2%) took one statin only and the remaining 28.8%, more than one. The most common, initially prescribed statin type was simvastatin (52.8%), followed by atorvastatin (38.5%); others included rosuvastatin, lovastatin, and fluvastatin, making up a total of five types. The distribution of initial statin prescriptions per year, showed an increasing trend, from 8,481 cases (9.9%) in 2010 to 13,807 cases (16.1%) in 2016.

Table 3

Statins prescription information
Characteristic of statins prescription	N	%
Number of statin prescriptions
1	28170	32.9
2 ~ 10	37965	44.3
>10	19613	22.9
Number of statin types
1	61060	71.2
> 1	24698	28.8
Type of first statin prescription
simvastatin	45280	52.8
atorvastatin	33051	38.5
rosuvastatin	5446	6.4
lovastatin	1166	1.4
fluvastatin	815	0.9
Year of first statin prescription
2010	8481	9.9
2011	9588	11.2
2012	13095	15.3
2013	13310	15.5
2014	13979	16.3
2015	13498	15.7
2016	13807	16.1

Twelve diagnostic categories were included in the study, among which the total number of AEs was 126. The top five categories of AEs were skin and subcutaneous tissue diseases, with 18 AEs (14.3%); circulatory system diseases, with 15 AEs (11.9%); digestive system diseases, with 13 AEs (10.3%); genitourinary system diseases, with 13 AEs (10.3%); and endocrine, nutritional, and metabolic diseases, with 12 AEs (9.5%). Based on published literature reviews on statin safety and systematic evaluative meta-analyses(1, 2, 11–14), this study identified 13 true-positive signals for statin-related AEs, which mainly included diabetes mellitus, as well as neurological, hepatic, muscular, and renal disorders.

Of the 85,758 patients in the statin group, 31,947 (37.3%) experienced at least one AE during the study period. In contrast, of the 138,429 study participants in the non-statin group, 39,761 (28.7%) had at least one AE. The number and incidence of AEs in each diagnostic category are shown in Table 4. The three categories with highest AE incidence in the statin group, involved skin and subcutaneous tissue disorders, musculoskeletal and connective tissue disorders, and ocular and adnexal disorders, with incidence rates of 7.6%, 7.1%, and 6.7%, respectively. Conversely, the three diagnostic categories with the highest incidence of AEs in the non-statin group were related to disorders of the skin and subcutaneous tissues, development and eruption of teeth, and the musculoskeletal system and connective tissues, with incidence rates of 7.2%, 6.8%, and 6.0%, respectively.

Table 4

Number and incidence of adverse events for each diagnostic category
ICD-10	diagnostic category	statin		non-statin
		N	%	N	%
D50-D89	Diseases of the blood and blood-forming organs and certain disorders involving the immune mechanism	718	0.8	1076	0.8
E00-E90	Endocrine, nutritional and metabolic diseases.	4510	5.3	4961	3.6
F00-F99	Mental, Behavioral and Neurodevelopmental disorders	188	0.2	358	0.3
G00-G99	Diseases of the nervous system	1602	1.9	1701	1.2
H00-H59	Diseases of the eye and adnexa	5759	6.7	7029	5.1
H60-H95	Diseases of the ear and mastoid process	3512	4.1	4486	3.2
I00-I99	Diseases of the circulatory system	2190	2.6	2321	1.7
J00-J99	Diseases of the respiratory system	1194	1.4	1585	1.1
K00-K93	Disorders of tooth development and eruption	5643	6.6	9360	6.8
L00-L99	Diseases of the skin and subcutaneous tissue	6484	7.6	9915	7.2
M00-M99	Diseases of the musculoskeletal system and connective tissue	6065	7.1	8324	6.0
N00-N99	Diseases of the genitourinary system	1792	2.1	3257	2.4

TreeScan mining signal results

A total of 29 positive signals (p < 0.05) were identified using TreeScan analysis. The corresponding ICD-10 code, Chinese diagnostic name (translated into English), number of observations, expected number, relative risk, and p-value of each signal, are listed in Table 5. Positive signals were compared with the standard signals to yield true positive signals, which are marked with asterisks in the table. Nine of the 29 positive signals were classified as true positives, namely, non-insulin-dependent and unspecified diabetes mellitus, respectively, other polyneuropathy, liver failure not classifiable elsewhere, other diseases of the liver, chronic kidney disease, unspecified and acute kidney failure, respectively, and other myopathies. Additional diagnostic categories for which false-positive signals were detected, included endocrine, ocular, circulatory, and genitourinary disorders, among others.

Table 5

TreeScan positive signal mining results
ICD	Diagnosis	Obs	Exp	RR	P
E03	Other hypothyroidism	82	71.6	1.15	0.001
E11	Type 2 diabetes mellitus^*	1385	1177.3	1.18	0.001
E14	Other specified diabetes mellitus^*	2399	2154.5	1.12	0.001
E16	Other disorders of pancreatic internal secretion	102	86.5	1.18	0.001
E87	Other disorders of fluid, electrolyte and acid-base balance	446	382.5	1.17	0.001
G44	Other headache syndromes	208	188.9	1.10	0.001
G62	Other and unspecified polyneuropathies^*	1261	1071.9	1.18	0.001
H11	Other disorders of conjunctiva	1224	1177.6	1.04	0.001
H26	Other cataract	1676	1480.2	1.14	0.001
H35	Other retinal disorders	289	253.1	1.14	0.001
H43	Disorders of vitreous body	580	534.5	1.09	0.001
H53	Visual disturbances	1890	1809.9	1.05	0.001
I42	Cardiomyopathy	88	72.7	1.21	0.001
I44	Atrioventricular and left bundle-branch block	35	26.7	1.30	0.001
I45	Other conduction disorders	30	22.0	1.34	0.001
I49	Other cardiac arrhythmias	1652	1394.0	1.19	0.001
J45	Asthma	967	927.5	1.04	0.001
K25	Gastric ulcer	1880	1788.4	1.05	0.001
K72	Hepatic failure, not elsewhere classified^*	143	127.1	1.13	0.001
K76	Other diseases of liver^*	1011	889.4	1.14	0.001
M10	Gout	541	515.3	1.05	0.001
M81	Osteoporosis without current pathological fracture	1920	1759.1	1.10	0.001
N18	Chronic kidney disease (CKD)^*	136	104.4	1.30	0.001
N19	Unspecified kidney failure^*	91	78.4	1.15	0.001
N17	Acute kidney failure ^*	10	8.7	1.15	0.002
G72	Other and unspecified myopathies ^*	9	7.9	1.14	0.002
I47	Paroxysmal tachycardia	117	106.5	1.10	0.002
N42	Other and unspecified disorders of prostate	128	117.2	1.09	0.002
N95	Menopausal and other perimenopausal disorders	325	309.0	1.05	0.015
* True positive signal

Evaluation of TreeScan results

Based on the detection results of the TreeScan method and the standard signal used to construct a contingency table (see Table 6), the sensitivity, specificity, PPV, NPV, correctness, Youden index, and AUC of the TreeScan method were calculated, to evaluate its signal detection ability. As shown in the table, the true and false positive signals of the TreeScan method were 9 and 20, respectively, whereas the false negative signal was 4, and true negative signal, 93. The formulas and calculated values of each evaluation index of the TreeScan method, are listed in Table 7. The results showed that the sensitivity, specificity, as well as positive and negative predictive values of TreeScan were 69%, 82%, 31%, and 96%, respectively. The overall evaluation index correctness, Youden index, and AUC were 81%, 52%, and 71%, respectively.

Table 6

Contingency table of TreeScan diagnostic test evaluation
Signal detection result	Standard signal		Total
	Positive events	Negative events
Positive signal	9	20	29
Negative signal	4	93	97
Total	13	113	126

Table 7

Evaluation index results of TreeScan method
Indicators	Formulas	Value (%)	95% CI
Sensitivity	a/(a + c)*100%	69.23	38.57–90.91
Specificity	d/(b + d)*100%	82.30	74.00-88.84
PPV	a/(a + b)*100%	31.03	20.81–43.52
NPV	c/(c + d)*100%	95.88	91.10-98.14
Accuracy	(a + d)/(a + b + c + d)*100%	80.95	73.00-87.40
Youden index	Sensitivity + Specificity − 1	51.53	12.50-79.75
AUC	-	75.77	62.24–89.29

TreeScan signal mining

The results of this study showed that TreeScan identified 29 positive signals from the raw data (p < 0.05), of which 9 were true positives when compared with the standard signals. In previous studies, researchers conducted only qualitative evaluations of results of the TreeScan method(7, 15–17). For example, Kulldorff et al. used the TreeScan method to uncover adverse reaction signals for two antifungal agents (terbinafine and itraconazole) and two thiazolidinedione hypoglycemic agents (pioglitazone and rosiglitazone)(7), resulting in 5 of the 10 identified signals being classified as adverse reactions, which included liver damage, allergic reactions, nausea, and vomiting. Of the remainder, four signals may have been generated by bias and confounding factors, and the remaining one may have been a potential adverse reaction, which required further validation.

In contrast, this study is the first to adopt a quantitative approach to evaluating the performance of the TreeScan method, using a diagnostic test evaluation method that determined the standard signal in advance and compared TreeScan-uncovered signals thereto. In the original data, the sensitivity, specificity, PPV, and NPV of the TreeScan method were 69%, 82%, 31%, and 96%, respectively, whereas the comprehensive index of correctness, Youden’s index, and AUC were 81%, 52%, and 71%, respectively.

Kulldorff et al. first applied the TreeScan approach to adverse drug reaction surveillance in 2013, in a study that mined the Health Maintenance Organization's (HMO's) health insurance database—which comprises data of 3 million participants—for adverse reaction signals(7). Their study obtained drug exposure data from pharmacy dispensing information, which included generic and trade names of drugs, dispensing dates, drug codes, and units, and dates of supply. The diagnostic information included ICD-9 diagnosis codes and dates, and in-/outpatient variables. Compared to the health insurance database, the Yinzhou regional medical database used in this study, collected richer information. In addition to key medication and diagnosis information, a wealth of demographic characteristics and lifestyle information, which could act as potential confounders, are available on this database. The Kulldorff study cohort was drawn from a sample of HMO health plan enrollees in nine of the United States, for the period from 1999 to 2003, comprising a total sample of nearly 3.41 million people. The present study cohort was selected from 1.3 million residents of Yinzhou District, Ningbo City, for the period from 2010 to 2016; ultimately, relevant criteria were met for the inclusion of approximately 220,000 study participants. Accordingly, the population sample size of this study was comparatively smaller and more limited in representativeness, which—to some extent—did not facilitate the detection of rare adverse reactions and affected the extrapolation of results.

Exposure and study outcomes

Active ADR surveillance using electronic medical databases, represents a new challenge in pharmacovigilance research. Electronic medical databases can provide relevant medical record data for large study populations at low cost and in a time-efficient manner, effectively compensating for the deficiencies of passive surveillance based on spontaneous submission(3). To delineate exposure factors, this study first defined patients with at least one statin prescription in their out-/inpatient records, as the statin group. This group was then further limited to patients who had had no statin prescription record for at least 1 year prior to their first statin prescription in the new user data, as well as in the matched data. Inclusion also required patients to have at least two statin prescription records within 6 months. Defining the statin group in this way helped reduce bias from prior medication use and misclassification(3, 18).

To define outcome events, this study focused on ICD-10 codes recorded in out-/inpatient diagnostic information, as potential adverse drug events. Unlike the WHO Adverse Reaction Terminology (WHO-ART) and International Medical Dictionary for Regulatory Activities (MedDRA)—two terminology sets used internationally for post-marketing ADR surveillance(19)—in which certain preferred terms are indicated for referral to ADRs, ICD-10 codes are mainly used for standardization of diagnosed diseases in electronic medical records. Some of these codes—such as gastric ulcer (K25), toxic liver disease (K71), and acute renal failure (N17)—can be used to document ADRs. A foreign systematic review summarized a set of ICD-10 codes for the identification of adverse drug events(10). That study was conducted by systematically searching five electronic databases for drug safety evaluation studies using ICD-10 codes, from which a total of 4,241 articles were retrieved and 41 articles ultimately included in the study. A total of 827 ICD-10 codes were summarized from these articles by clinical experts and coders, to identify adverse drug events. The present study referred to that set of codes and intercepted the first three digits of relevant codes to compile 126 standard signals for adverse drug events, covering 12 diagnostic categories. Outcome events in this study were thus defined as 126 ICD-10 codes recorded in out-/inpatient diagnoses, within 180 days of prescription. The 180-day time window was mainly selected based on foreign ADR signal mining methodological studies; setting a time window represents an effective method, especially when the start and end dates of medication administration for patients are not accurately available(20, 21).

TreeScan can be applied as a signal detection method in Chinese regional healthcare databases for drug safety surveillance, simultaneously evaluating small and large numbers of potential AEs (subsets of diagnostic codes). It can also adjust to accommodate multiple testing approaches that are inherent when many overlapping groups are evaluated. In future, a comparative evaluation study should be conducted to further evaluate its performance and explore its proper application conditions.

Acknowledgments

We thank the Health and Family Planning Bureau of Yinzhou District for authorizing use of the EMRs, and the Centers for Disease Control, for providing access to the databases used in the study.

Author Contributions

HLi and SZ contributed to the study conception and design. HLi and HZ analyzed the data and drafted the manuscript. HLi, HZ, and SZ critically revised the manuscript. All authors were involved in the interpretation of data and have read and approved the final manuscript.

Data Availability Statement

The data analyzed in this study are subject to the following restrictions: data can only be used locally. Requests to access these datasets should be directed to Hongbo Lin at [email protected]

Competing interests

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as potential conflicts of interest.

Ethics Statement

Studies involving human participants were reviewed and approved by the Peking University Health Science Center’s Institutional Review Board. Written informed consent for participation was not required for this study, in accordance with national legislation and institutional requirements.

Funding

This study was supported by the National Natural Science Foundation of China (grant numbers 81973146 and 72004151).

Adhyaru BB, Jacobson TA. Safety and efficacy of statin therapy. Nat Rev Cardiol (2018) 15(12):757–69. Epub 2018/10/31. doi: 10.1038/s41569-018-0098-5. PubMed PMID: 30375494.
Newman CB, Preiss D, Tobert JA, Jacobson TA, Page RL, 2nd, Goldstein LB, et al. Statin Safety and Associated Adverse Events: A Scientific Statement From the American Heart Association. Arterioscler Thromb Vasc Biol (2019) 39(2):e38-e81. Epub 2018/12/26. doi:
10.1161/ATV.0000000000000073. PubMed PMID: 30580575.
Wadman M. Experts call for active surveillance of drug safety. Nature (2007) 446(7134):358-9. Epub 2007/03/23. doi: 10.1038/446358b. PubMed PMID: 17377551.
Ghosh P, Dewanji A. Effect of reporting bias in the analysis of spontaneous reporting data. Pharm Stat (2015) 14(1):20–5. Epub 2014/11/08. doi: 10.1002/pst.1657. PubMed PMID: 25376637.
Huang YL, Moon J, Segal JB. A comparison of active adverse event surveillance systems worldwide. Drug Saf (2014) 37(8):581–96. Epub 2014/07/16. doi: 10.1007/s40264-014-0194-3. PubMed PMID: 25022829; PubMed Central PMCID: PMCPMC4134479.
Yang Y, Zhou X, Gao S, Lin H, Xie Y, Feng Y, et al. Evaluation of Electronic Healthcare Databases for Post-Marketing Drug Safety Surveillance and Pharmacoepidemiology in China. Drug Saf (2018) 41(1):125 – 37. Epub 2017/08/18. doi: 10.1007/s40264-017-0589-z. PubMed PMID: 28815480.
Kulldorff M, Dashevsky I, Avery TR, Chan AK, Davis RL, Graham D, et al. Drug safety data mining with a tree-based scan statistic. Pharmacoepidemiol Drug Saf (2013) 22(5):517–23. Epub 2013/03/21. doi: 10.1002/pds.3423. PubMed PMID: 23512870.
Lin H, Tang X, Shen P, Zhang D, Wu J, Zhang J, et al. Using big data to improve cardiovascular care and outcomes in China: a protocol for the CHinese Electronic health Records Research in Yinzhou (CHERRY) Study. BMJ open (2018) 8(2):e019698. Epub 2018/02/15. doi: 10.1136/bmjopen-2017-019698. PubMed PMID: 29440217; PubMed Central PMCID: PMCPMC5829949.
Coloma PM, Avillach P, Salvo F, Schuemie MJ, Ferrajolo C, Pariente A, et al. A reference standard for evaluation of methods for drug safety signal detection using electronic healthcare record databases. Drug Saf (2013) 36(1):13–23. Epub 2013/01/15. doi: 10.1007/s40264-012-0002-x. PubMed PMID: 23315292.
Hohl CM, Karpov A, Reddekopp L, Doyle-Waters M, Stausberg J. ICD-10 codes used to identify adverse drug events in administrative data: a systematic review. J Am Med Inform Assoc (2014) 21(3):547 – 57. Epub 2013/11/14. doi: 10.1136/amiajnl-2013-002116. PubMed PMID: 24222671; PubMed Central PMCID: PMCPMC3994866.
Law M, Rudnicka AR. Statin safety: a systematic review. Am J Cardiol (2006) 97(8A):52C-60C. Epub 2006/04/04. doi: 10.1016/j.amjcard.2005.12.010. PubMed PMID: 16581329.
Sattar N, Preiss D, Murray HM, Welsh P, Buckley BM, de Craen AJ, et al. Statins and risk of incident diabetes: a collaborative meta-analysis of randomised statin trials. Lancet (2010) 375(9716):735–42. Epub 2010/02/20. doi: 10.1016/S0140-6736(09)61965-6. PubMed PMID: 20167359.
Iwere RB, Hewitt J. Myopathy in older people receiving statin therapy: a systematic review and meta-analysis. Br J Clin Pharmacol (2015) 80(3):363–71. Epub 2015/06/03. doi: 10.1111/bcp.12687. PubMed PMID: 26032930; PubMed Central PMCID: PMCPMC4574822.
Mancini GB, Baker S, Bergeron J, Fitchett D, Frohlich J, Genest J, et al. Diagnosis, Prevention, and Management of Statin Adverse Effects and Intolerance: Canadian Consensus Working Group Update (2016). Can J Cardiol (2016) 32(7 Suppl):S35-65. Epub 2016/06/28. doi: 10.1016/j.cjca.2016.01.003. PubMed PMID: 27342697.
Wintzell V, Svanstrom H, Melbye M, Ludvigsson JF, Pasternak B, Kulldorff M. Data Mining for Adverse Events of Tumor Necrosis Factor-Alpha Inhibitors in Pediatric Patients: Tree-Based Scan Statistic Analyses of Danish Nationwide Health Data. Clin Drug Investig (2020) 40(12):1147–54. Epub 2020/10/27. doi: 10.1007/s40261-020-00977-5. PubMed PMID: 33104987; PubMed Central PMCID: PMCPMC7701063.
Huybrechts KF, Kulldorff M, Hernandez-Diaz S, Bateman BT, Zhu Y, Mogun H, et al. Active Surveillance of the Safety of Medications Used During Pregnancy. Am J Epidemiol (2021) 190(6):1159–68. Epub 2021/01/11. doi: 10.1093/aje/kwaa288. PubMed PMID: 33423046.
Fralick M, Kulldorff M, Redelmeier D, Wang SV, Vine S, Schneeweiss S, et al. A novel data mining application to detect safety signals for newly approved medications in routine care of patients with diabetes. Endocrinol Diabetes Metab (2021) 4(3):e00237. Epub 2021/07/20. doi: 10.1002/edm2.237. PubMed PMID: 34277962; PubMed Central PMCID: PMCPMC8279599.
Hempenius M, Groenwold RHH, de Boer A, Klungel OH, Gardarsdottir H. Drug exposure misclassification in pharmacoepidemiology: Sources and relative impact. Pharmacoepidemiol Drug Saf (2021) 30(12):1703–15. Epub 2021/08/17. doi: 10.1002/pds.5346. PubMed PMID: 34396634.
Brown EG. Effects of coding dictionary on signal generation: a consideration of use of MedDRA compared with WHO-ART. Drug Saf (2002) 25(6):445–52. Epub 2002/06/20. doi: 10.2165/00002018-200225060-00009. PubMed PMID: 12071782.
Ryan PB, Madigan D, Stang PE, Overhage JM, Racoosin JA, Hartzema AG. Empirical assessment of methods for risk identification in healthcare data: results from the experiments of the Observational Medical Outcomes Partnership. Stat Med (2012) 31(30):4401–15. Epub 2012/09/28. doi: 10.1002/sim.5620. PubMed PMID: 23015364.
Ryan PB, Stang PE, Overhage JM, Suchard MA, Hartzema AG, DuMouchel W, et al. A comparison of the empirical performance of methods for a risk identification system. Drug Saf (2013) 36 Suppl 1:S143-58. Epub 2013/11/06. doi: 10.1007/s40264-013-0108-9. PubMed PMID: 24166231.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Detecting adverse reaction signals to statins in a Chinese regional healthcare database, using a tree-based scan statistic method

Status:

Version 1

Abstract

Figures

Introduction

Methods

Statistical analysis

Results

Discussion

Conclusions

Declarations

References

Additional Declarations

Status:

Version 1