Adverse drug reactions (ADRs) are considered a major international clinical concern and a significant cause of death in medical facilities [1]. It recommended that safer medical systems be developed to enable the early discovery of drug side-effects [23]. Although a voluntary reporting policy has been implemented to reduce ADRs, relatively few side-effects of drugs have been reported when compared to the actual number of incidents of adverse drug effects [4].

In the past, the occurrence of ADRs after the use of prescription drugs could be detected when the patients and doctors noticed adverse symptoms. Recently, many studies have attempted to detect ADRs using Electronic Medical Record (EMR) data [5]. Since the conversion of conventional paper-based health record systems to EMR systems, a considerable amount of medical information has been collected. Accordingly, diverse efforts have continued to systematically identify adverse drug effects without depending on voluntary reports. The large, computerized clinical databases associated with EMR can be used with various ADR detection methods. Moreover, they can serve as a source of information that clinicians can use to readily and appropriately respond in the event a patient suffers from an ADR [6].

In this study, we created an algorithm to enable early detection of ADRs using EMR data. We selected a 3-hydroxy-3-methylglutaryl-coenzyme A reductase (HMG-CoA reductase) inhibitor (HMG-CoA reductase inhibitor, which is a statin). A statin is a representative drug for the treatment of hyperlipidemia [78]. It facilitates both the primary and secondary prevention of cardiovascular disease by lowering low-density lipoprotein cholesterol [9]. However, statins cause various ADRs, such as myopathy, hepatotoxicity, nephrotoxicity, and insulin resistance disorder [1011]. Statin ADRs should be considered on a case-by-case basis because they are distinguished according to their specific metabolic pathways or drug interactions.

To improve the safety of patients who use statins, we developed an algorithm for automatically detecting liver injury due to statin use from EMR data, including coexisting diseases, laboratory findings, and concomitant drugs. To verify the algorithm's performance, a primary ADR analysis of the algorithm was performed after extraction of data of patients who were prescribed statins. A secondary chart review analysis was conducted. After comparing the primary algorithm assessment and secondary chart review analyses, we identified approaches to improving the algorithm's performance.

Statin-induced liver injury increases the AST or ALT level. Its incidence is reported to be 0.5%–2.0% [151617]. Some studies have reported that statin use does not lead to liver damage and that there is no correlation between ALT increase [1518] and statin use. Moreover, it has been reported that any increase in ALT is temporary and reversible [1920]. That is, there is liver damage, but it is not severe, according to recent reports. However, some controversy still exists regarding the effects of statins. Current guidelines demand caution when ALT increases after statin use. In Korea, the guidelines state that when the AST or ALT level in a patient reaches thrice its normal value, statin use needs to stop. For Asian patients, it is understood that even more special caution is recommended while using statins. In Korea, prior to the first-time use of statin, baseline AST or ALT is measured, and after statin use begins, re-checking at 6 weeks and 12 weeks is recommended. In general, statin-induced liver injury is naturally reversed and normalized if the patient discontinues the use of statins. However, in rare cases, idiopathic acute liver failure, which is associated with statin use, causes irreversible liver damage, eventually requiring a liver transplant or causing early death [2122]. Therefore, detecting ADRs at an early stage is necessary to prevent statininduced liver injury from causing serious damage. Because patient safety is of critical importance, detection of an ADR at an early stage is crucial.

EMR data include data for detecting the occurrence of ADRs, such as the time of drug prescription, symptom appearance, and detailed clinical progress. Therefore, EMR data may be a source of information for detecting potential ADRs. In addition, screening specific laboratory findings by using EMR data enables the continuous monitoring of a large number of patients, while using fewer data resources than a chart review [2324]. The Sentinel Project of the US Food and Drug Administration [25] analyzes ADR signals from multi-site EMRs. In addition, the Korean Comparison of the Laboratory Extreme Abnormality Ratio (CLEAR) algorithm analyzes ADR using laboratory findings [26]. However, neither can accurately detect the specific ADR of a particular drug for an individual patient because these studies were developed to detect a number of ADRs for large-scale drugs. Thus, in terms of individual patient safety, both a scientific methodology and proof that an individual specific ADR can be detected via ordinary monitoring using EMRs are needed.

Automatic ADR assessment through the proposed algorithm can detect drug-induced ADR at early stages. As determined by the algorithm, the occurrence rate of liver injury was 1.4%. The PPV-associated accuracy was 73.8%, which was somewhat low. The reasons for this low result were that the diseases coexisting with liver disease were not completely filtered by the algorithm, and not all of the concomitant drugs that cause hepatotoxicity were considered by the algorithm [2728]. This may be because information about a rechallenge of statins specified by the WHO-UMC standard was not included in the algorithm. The current algorithm is very simple and will be enhanced in subsequent work. If the current algorithm is revised to exclude any coexisting liver diseases and include the remaining concomitant drugs, the performance of the algorithm will improve. Moreover, liver injury due to statin use could be assessed with greater speed and accuracy. Meanwhile, the algorithm's rapid automatic ADR assessment can replace manual ADR assessments. The algorithm can assess ADRs with a blood test in a shorter amount of time. During treatment, doctors check important items that cannot be automatically assessed by the system. Thus, the time required to confirm the occurrence of an ADR will effectively decrease.

Our algorithm includes concomitant drug information, diagnosis information, and clinical test information to decrease errors in the ADR assessment. Because we identified improvements for the algorithm through the results of the chart review, the core elements of this algorithm are coexisting diseases, lab test results, and concomitant drugs. It should be noted that the CLEAR algorithm has some limitations because it checks only for abnormalities of laboratory tests to detect ADRs [26]. Furthermore, in the study of Sai et al. [29] of the development of an algorithm to detect statin-induced myopathy, ADRs were assessed by medical practice, adjustments of the times of ADR events, diagnoses, and laboratory data. However, medical practice and adjustments of the times of ADR events are limited by the requirement of frequent blood tests; thus, we could not apply this method to our algorithm based on the results from outpatients visiting every three months. Therefore, we included concomitant drug information, diagnostic information, and laboratory test results. We compared the prescription dates of the concomitant drugs with the occurrence dates of ADRs. To confirm the effects of the concomitant drug on increases in liver enzyme levels, we checked whether the liver enzyme levels were normalized after the patients stopped using the concomitant drugs. Based on this work, our study can be directly applied in real practice and can be useful when upgrading the algorithm later because there are diverse data in EMRs. Beside ALT or ALP, additional necessary laboratory items can be added.

Despite the above contributions, this study had some limitations. When applying the data extracted from EMRs to the developed ADR assessment algorithm, we found that almost 40% of the patients did not have the necessary information. Therefore, the EMR structure should be improved so that related tests can be performed before a drug is prescribed, and the results of the blood tests can be automatically recorded [26]. Furthermore, this research did not access real clinical ADRs that were not discovered by the actual algorithm. In other words, sensitivity and specificity were not identified. However, there was no significant difference between the incidence rate of statin-induced liver injury in many previous studies and the incidence rate of statin-induced liver injury estimated by the algorithm in this study. Moreover, the algorithm was developed without considering that patients may have stopped using statins or changed the type of statin they were using. These aspects will be addressed in subsequent work. Additionally, because coexisting diseases were recorded as free descriptions in the outpatient record, the analysis was not speedy. This information was only found through the chart review. Therefore, these data must be recorded not as text but with corresponding codes in the EMR. Thus, improvements in EMR systems are necessary to enable easy extraction of data [2630]. Many weak points regarding the proof methods of this research exist, but they will be continuously addressed and backed up in future studies; we hope this research will be considered as the first step in this process.

It is desirable to establish a gold standard regarding adverse drug effects by detecting clues of ADR while monitoring side effects and devising countermeasures. However, for drugs of which side effects are already known, if drugs that can prevent these side effects can be preferentially prescribed, continuous monitoring should be systematically possible based on indices of side effects. With such a method, the time required for doctors to prescribe drugs and the length of time during which patients suffer side effects should be minimized. In our study, to detect early statin-induced liver injury, we developed an automatic ADR assessment algorithm that uses the records of coexisting diseases, laboratory tests, and concomitant drugs. Nevertheless, the current system algorithm that we created is too simple. We expect that a more reliable algorithm can be created if we conduct and supplement clinical studies in terms of adverse drug effects. The performance of the algorithm is expected to be improved through larger studies, although the algorithm cannot perfectly detect hepatotoxicity. In addition, we look forward to the clinical application of this algorithm after further validation.