This study was performed in accordance with the World Medical Association Declaration of Helsinki, and the study protocol was approved by the institutional review board of the Osaka University Hospital (OUH), National Cerebral and Cardiovascular Center (NCVC), and Tottori University Hospital (TUH). This study was an observational study and did not obtain individual informed consent from the participants included in the study. However, the study protocol was posted on our webpage, giving the study participants an opportunity to opt out.

Because each CPOE system has its own database, the systems have different structures. We first developed an intermediate database to unify the database structure. The data from the original CPOE database were transferred to this intermediate database. We then generated the medication history by applying a medication history generation (MHG) program to the intermediate database (Figure 1). The medication history generation program was developed with Microsoft Visual Basic for Applications 7.0, and Microsoft Access 2010 was used for the intermediate database. Both the program and the database were installed on a laptop PC ( Intel Core i7-2640M CPU; 8 GB of memory) with the Microsoft Windows operating system.

The medication history includes the start and the end dates of medication for each medicine prescribed to a patient. To construct the medication history database, the CPOE records were combined with consideration to the remaining medicine (Figure 2).

First, the CPOE records for each medicine taken by an individual patient were extracted and combined sequentially from the oldest record to the newest record. As shown in Figure 2A, in cases where the last day of prescription 1 (P-1) was after the first day of P-2, the first day of P-1 was set as the start date of medication while the last day of P-2 was set as the end date, and the amount of the remaining medicine was estimated. In the case that a gap period lay between the last day of P-2 and the first day of P-3, if the amount of the remaining medicine was not less than the amount of medicine that would have been consumed during the gap period, P-2 and P-3 were combined.

We estimated the amount of the remaining medicines due to noncompliance by the patient, assuming that the rate of missed doses was constant. Accordingly, we set an unused medicine index (UMI), which indicated the rate of missing doses as a ratio of the period in which patients actually took the medicine to the prescription period. The amount of remaining medicine due to missing doses in P-4 was calculated (Figure 2B). If the amount of the remaining medicine was greater than that needed for the days during the gap period, P-4 and P-5 were combined. Figure 2 shows the algorithm for generating the medication history database.

The appropriate UMI value was determined by generating the medication history records for 9 medicines (pravastatin, cilostazol, isosorbide, nifedipine, ursodeoxycholic acid, rebamipide, amlodipine, aspirin, and methylcobalamin), which were considered to be long-term prescriptions. We set the UMI value as 1.4 because unnatural short-term gap periods tended to be observed when the UMI was <1.3 and because gap periods of a few months (considered to be the cessation of medicine) tended to be combined when the UMI was >1.5. To evaluate the validity of the UMI, we randomly selected 9 patients who had been treated for cardiovascular disease for >5 years and generated 725 medication history records. The medication history records were reviewed by the chief physicians, resulting in 98% of the records being considered appropriate.

We detected the occurrence of ADEs based on the chronological relationship between the presence of the adverse event and the medication period. In this study, we focused on identifying ADEs with ALT elevation, which is known to reflect hepatocellular injury-type DILI. The elevation of ALT was selected because, in the RUCAM, the severity of hepatocellular injury-type DILI is defined by serum ALT. The ALT values were obtained from the laboratory test data in the EMR database. The criteria for the diagnosis of ADE with ALT elevation are shown in Table 1. ADEs with ALT elevation were detected during the medication period, and those with a decrease in the ALT level were detected after the cessation of the medication (criteria 1 and 2). If the elevated value decreased during the medication period, then the medicine was considered not to be causative; thus, it was excluded as a cause of ADE (criterion 3). Because it was difficult to distinguish an ALT elevation caused by a previous liver injury, viral hepatitis, or an operation from hepatocellular injury-type DILI, we excluded patients with any of these factors (criterion 4-III).

In the present study, EMR data were obtained from 3 medical facilities: OUH, NCVC, and TUH. These medical institutions have independent EMR systems. In the study period, the data from a total of 1,587,939 patients were registered, and the total number of CPOE records was 37,935,783 (an average of 23.9 records per patient; Table 2).

The target medicines were aspirin, clopidogrel, and ticlopidine. These are antiplatelet drugs that have been reported to cause hepatocellular injury-type DILI [26-28]. Earlier studies have suggested that clopidogrel is associated with a lower risk of hepatocellular injury-type DILI in comparison to ticlopidine [29].

To calculate the rates of ADE with ALT elevation that occurred with each medicine, we counted the number of patients who met the diagnostic criteria (Table 1). The severity of ADE with ALT elevation was categorized according to the maximum ALT value as mild elevation (maximum ALT ≥40 IU/L), moderate elevation (maximum ALT ≥80 IU/L), and severe elevation (maximum ALT ≥200 IU/L). The rate of ADEs with ALT elevation was calculated by dividing the number of ADE patients by the number of patients who took the targeted medicine, and the ALT values were tested at least 3 times (before, during, and after the medication period).

We selected the patients with moderate and severe ALT elevation (maximum ALT ≥80 IU/L) whose medical records were recorded electronically at OUH and TUH and checked the progress notes recorded from 3 days before to 3 days after the date of the peak ALT value. The numbers of medical records subjected to review for each of the drugs were as follows: aspirin (n=83), clopidogrel (n=29), and ticlopidine (n=8). These records were used to determine whether or not the elevation of ALT was due to an ADE. The ADE cases were categorized into 3 groups: (1) ADE caused by the targeted medicine, (2) ADE caused by a concomitant medicine, and (3) offending medicine not identified.

The RUCAM is the standard method for detecting DILI. The RUCAM uses a 5-stage scoring system to assess the possibility of DILI by classifying the condition as hepatocellular, cholestatic, or mixed based on the laboratory test data and clinical data.

We compared the accuracy of detecting hepatocellular-type ADE between our method and the RUCAM. Patients with ALT levels of >200 were included in the analysis (10,608 patients from OUH and 5464 patients from TUH).

The primary screening was performed to select hepatocellular-type ADE for the RUCAM. The screening criterion was as follows: ALT level >200 and (ALT/upper limit of normal/(alkaline phosphatase/upper limit of normal)>5 within 90 days of the first day of using the verified medication. Next, we determined the RUCAM score based on a review of medical records. Probable and highly probable scores according to the RUCAM system were classified as hepatocellular-type DILI in this study.

Multiple comparisons were performed using the Ryan method, and the Fisher exact test was used to compare the rates of ADE. P values of <0.05 were considered to indicate statistical significance. All statistical analyses were performed using the R software version 3.1.2 (The R Foundation for Statistical Computing).

Accurate demonstration of the start and end dates of a medication period is important in pharmacoepidemiologic research. However, the CPOE records in EMRs cannot clearly demonstrate the total duration of the medication period. In the present study, we generated a medication history database from the CPOE databases of 3 hospitals and systematically diagnosed ADEs with ALT elevation according to the chronological relationship between the changes in ALT values and the duration of medication using a medication history database. Because the medication history database can be applied not only to the detection of ADEs but also to crossover studies that compare drug efficacy in the same patients, it can become a basis for pharmacoepidemiologic research.

The comparison of the RUCAM and our method revealed that the rates of ALT and ALP testing influenced the accuracy of the RUCAM in the detection of ADEs. In a prospective study, laboratory test data can be obtained according to a research plan. However, in a retrospective study, missing data often become problematic. Scoring in the RUCAM requires information such as the use of concomitant medications, drug risk information, the presence or absence of a rechallenge, and the history of alcohol consumption. This information is not registered as structured data in EMRs. In this study, a review of medical records was needed to determine the score for the RUCAM. In contrast, our method used only standardized data, such as laboratory test data, prescription data, disease name data, and surgical data. For this reason, our method is applicable to the detection of ADEs in a retrospective analysis of big data generated by EMRs.

The population characteristics greatly affect the rate of adverse events. In clinical trials, the incidence of adverse events may be accurate because blood testing is routinely performed in all patients. On the other hand, in observational studies, the timing of blood testing differs for each patient. There may be great differences in the rates of adverse events depending on how the study population is defined. A previous clinical study in Japan reported that the rates of serious liver injury among patients receiving ticlopidine and clopidogrel were 13.6% (129/948) and 5.1% (115/2261), respectively [30,31]. However, these studies had different study populations, and caution must be exercised when interpreting the comparison of the rates of adverse events. The present method determined the rates of adverse events for some medicines under the same conditions for ticlopidine (188/5076 ,3.70%) and clopidogrel (226/3974, 5.69%); thus, this method could be used to compare the risk of adverse events between medicines (ticlopidine therapy is associated with a greater risk of developing ADEs in comparison to clopidogrel).

When physicians suspect an ADE with ALT elevation, all of the medicines that might have caused the ADE are likely to be discontinued. Thus, it was difficult to differentiate the causative medicine from the concomitant medicines using our method. Our method demonstrated the maximum rate of ADEs with ALT elevation induced by a targeted medicine, assuming that the targeted medicine was the causative medicine in all cases.

Although aspirin has been reported as a cause of liver injury, the rate in Asian populations remains unclear. According to the clopidogrel versus aspirin in patients at risk of ischemic events (CAPRIE) Steering Committee report, the rates of liver injury in patients receiving aspirin and clopidogrel were 2.97% (285/9599) and 3.15% (302/9586), respectively, which are in line with the rates obtained in the present study (aspirin: 868/26,059, 3.33%; clopidogre: 188/5076, 3.70%) [32]. The rates of severe liver injury in the same report were 0.19% for aspirin (93/5076, 0.36% in this study) and 0.11% for clopidogrel (22/5076, 0.43% in this study). Similar to our study, the rates of severe liver injury did not differ between patients using aspirin and those using clopidogrel.

Even though the absolute risk of a medicine is difficult to estimate, our method can estimate the upper limit of the risk. Furthermore, for some medicines, our method can estimate the risk of for ADE with ALT elevation one at a time under the same conditions, and the risk can be compared among different medicines.

In this study, we used the medication history database created from CPOE records to detect DILI, but we did not detect all cases of DILI. First, we focused on elevated serum ALT levels. Elevated serum ALT can capture hepatocellular-type DILI, but it may not detect cholestatic-type DILI, which is characterized by elevation of the serum ALP level. Second, we were not able to detect DILI that did not meet our diagnostic criteria, such as delayed DILI, even the hepatocellular-type DILI. This type of detection requires a different set of criteria.

The generation of a medication history database enabled us to detect ADEs with ALT elevation through the chronological relationship between the medication period and occurrence of liver injury. As our method used only standardized data from EMRs, it was possible to analyze real-world data accumulated by EMRs in multiple hospitals. Although our method could not identify the causative medicine among concomitant medicines, it was possible to compare the risk of ADEs for different medicines.