Background
QTc-prolongation is known as a risk factor for developing ventricular arrhythmias such as Torsade de Pointes (TdP), which may eventually lead to sudden cardiac death. Therefore, a prolonged heart-rate corrected QT(c) interval is used as electrocardiogram (ECG) marker for an increased risk of TdP; and thus a prolonged QTc-interval should be avoided in patient care as a part of risk minimization [
1‐
3].
QTc-prolongation is defined as a QTc-interval > 450 ms in males and > 470 ms in females according to the European Medicine Agency guidelines [
4,
5]. However, arrhythmias are frequently associated with QTc-intervals exceeding 500 ms. [
6‐
8] A prolonged QTc-interval often represents a delayed ventricular repolarization. Roden et al. introduced a theory where some physiological mechanisms create a buffer to maintain normal ventricular repolarization, the so-called repolarization reserve. Several risk factors and genetic predisposition can reduce this repolarization reserve causing abnormalities in the ventricular repolarization [
9,
10]. Consequently, multiple risk factors are frequently present in case reports describing patients who developed serious QTc-prolongation or TdP [
11,
12].
Several drugs are also responsible for developing QTc-prolongation known as drug-induced QTc-prolongation. Currently, over 190 drugs are associated with QTc-prolongation according to the CredibleMeds® QT drug lists of the Arizona Center for Education and Research on Therapeutics (AZCERT). AZCERT categorizes QTc-prolonging drugs into three categories representing the level of certainty on the risk of TdP. More than 50 drugs are categorized as drugs with a known risk of TdP [
13]. Many of these drugs such as antibiotics and antidepressants are widely used in clinical practice. QTc-prolonging drugs are not further classified with respect to the extent of QTc-prolongation. Also, the exact risk of developing QTc-prolongation when using a combination of QTc-prolonging drugs is unknown. For healthcare professionals, such as physicians and pharmacists, it is difficult to decide whether or not it is safe to proceed treating a patient with combinations of two or more QTc-prolonging drugs, and in whom additional checks of ECGs after treatment initiation are needed.
Other risk factors include hypokalemia, hypomagnesemia, heart diseases (i.e. ischemic heart diseases, heart failure, and arrhythmia such as atrial fibrillation), and renal impairment. Also, demographic risk factors such as an older age, female sex and genetic predisposition are associated with QTc-prolongation [
2,
12,
14‐
16]. However, the impact of these risk factors on the extent of QTc-prolongation is largely unknown, which makes it challenging to identify patients at risk for QTc-prolongation.
In the Netherlands, QT-DDI alerts are generated by the Computerized Physician Order Entry (CPOE) systems when two or more QTc-prolonging drugs with a known risk of TdP are combined. QT-DDI alerts are generated according to the so-called ‘G-Standard’, a Dutch drug database which supports the different processes in healthcare, such as prescription, dispensing, ordering, reimbursement, and decision support [
17]. The current guidelines incorporated in the ‘G-Standard’ regarding QT-DDIs suggest to substitute or remove one of the interacting agents or perform routine ECG monitoring. As a result, first-line treatments are frequently not adhered to when one of the interacting agents is substituted, especially in primary care where ECG monitoring is often not feasible. In tertiary care, low adherence to guidelines result in many overridden DDI alerts by physicians [
18]; and ECG monitoring is rarely performed when QT-DDI alerts are overridden [
19,
20]. With the increasing number of QTc-prolonging drugs, QT-DDI alerts will reduce the physician responsiveness to this particular type of alert, also known as
alert fatigue. The use of a smart algorithm which generates specific alerts will reduce alert fatigue in clinical practice.
Discussion
We have developed a tool which enables the identification of patients with an increased risk of QTc-prolongation when using two or more QTc-prolonging drugs with a known risk of TdP. We chose to develop a tool based on seven predictors, that could easily be implemented in everyday practice. The model was externally validated using an independent dataset of a general teaching hospital, showing the robustness of the model. Implementing such a model in clinical practice might enhance the identification of high-risk patients which will support healthcare providers in selecting patients in whom the risk of QTc-prolongation is such that therapy adjustment and/or additional ECG monitoring is required. At the same time such a model might also identify patients at low risk for developing cardiac arrhythmia, and in whom there is no need for monitoring ECGs after drug initiation, but further improvement of the tool is needed. However, it should be noted that the tool is not designed to fully replace healthcare providers in handling QT drug-drug interactions.
The model, developed in a university medical center, showed insufficient discrimination abilities (AUROC < 0.60) when applied to a dataset from a general teaching hospital. In the development cohort, we used a cut-off value of 450 ms in men and 470 ms in women for a prolonged QTc-interval using the Bazett formula. The Bazett formula often overestimates the QTc-interval in patients with sinus tachycardia [
37]. In the development cohort, 91.6% of the patients had heart rates within the range of 60–100 bpm. Arrhythmias are often associated with QTc-intervals exceeding 500 ms [
1,
8,
38]. Therefore, we performed a post hoc analysis to compare both reference values. The AUROC curve was 0.54 (95% CI 0.51–0.56) for QTc > 450/470 ms, but increased when QTc-prolongation was defined as QTc > 500 ms (0.59, 95% CI 0.54–0.63). We were aiming for high sensitivities to generate low numbers of false negatives, in order to not miss patients at high risk for TdP. The model was more sensitive in identifying QTc-intervals exceeding 500 ms. The optimized cut-off value of 6 resulted in a sensitivity of 76.6% for the prediction of QTc > 450/470 ms and 83.9% for the prediction of QTc > 500 ms.. However, the low specificity (27.5%) means that the model incorrectly labels patients at risk for QTc-prolongation. Nevertheless, we focused on optimizing sensitivity in order to prevent missing patients at risk for QTc-prolongation by accepting sub-optimal specificity values. As the current guidelines generate alerts in all patients, the guidelines lead to a specificity of zero, so even a specificity of 27.5% is an improvement. A perfect prediction model is not feasible because there is a wide variability in the QTc-interval independent of risk factors. And also, the incidence of QTc-intervals above 500 ms is relatively low [
8,
19,
39]. Nevertheless, the sensitivity and specificity value of 83.9 and 27.5% should be optimized before broad implementation in clinical practice can be recommended.
Ideally, the model should be developed and validated with TdP as primary outcome. As linear correlation is lacking, it is questionable whether a prolonged QTc-interval is an adequate marker for predicting the risk of TdP [
40]. Unfortunately, it is nearly impossible to identify cases of TdP, because ECGs are frequently not available to ensure TdP actually occurred. Furthermore, even in high risk populations the incidence of TdP is extremely low, so exceptionally large patient populations are needed to study TdP as primary endpoint. So a prolonged QTc-interval is still the most validated and frequently used surrogate marker in clinical practice [
2,
41].
Several studies have already introduced risk models for predicting QTc-prolongation/TdP. Haugaa et al. developed the ‘pro-QTc’ risk score, however, the primary endpoint in their study was mortality which is a different endpoint than the primary endpoint used in this study [
38]. Tisdale et al. developed a risk score via a similar approach, but included only patients admitted to cardiac care units [
15]. Consequently, generalizability to a general population may be limited. Vandael et al. recently developed an optimized RISQ-PATH score to detect high-risk patients for developing QTc-prolongation [
42]. However, when this model was applied to patients using two or more QTc-prolonging drugs, the sensitivity of the model was 94.5%, but the specificity of the model was even lower than our model (22.1%). Moreover, the RISQ-PATH score of Vandael et al. consists of too many predictors which are frequently not available and, therefore, this tool cannot be used in clinical practice. In addition, this tool needs to be implemented in the clinical decision support system before it is applicable in primary care, which does not seem feasible with the current electronic patient health records. We aimed to develop a risk score to detect high-risk patients when using two or more QTc-prolonging drugs which is easily applicable in both primary and hospital care.
A major strength of this study is that we externally validated the risk model in an independent dataset from a general teaching hospital. External validations are able to determine the generalizability of predicting models in different settings [
43].
Several limitations of our study need to be addressed. First, the study was limited by a single-center design for model development; however, patients were admitted to all general nursing departments representing a general hospital population. Second, the sample size was relatively small which increased the risk of model overfitting; a common problem in models derived from small datasets. By validating the model in a large external dataset and by adding predictors based on a review of literature, the risk of overfitting was minimized [
25,
26]. In the validation cohort, there might have been selection bias as the prevalence of QTc-prolongation (41.5%) was quite high compared to the overall prevalence found in the literature review (21%). We retrospectively collected these data, so presumably, ECGs were mainly recorded in high-risk patients. Our model does not take into account the QTc-interval at baseline. Given that the risk on QTc-prolongation increases when a high baseline QTc-interval is present, we chose to exclude this potential predictor because baseline ECGs are frequently not available in clinical practice. Also, the small dataset precluded the inclusion of too many predictors in the model. Third, the tool does not take into account the variety of QT-DDIs as our aim was to develop an easily obtainable model that can be used in different healthcare settings. Because of the different pharmacological pathways of the QTc-prolonging drugs via inhibition of the hERG channels or Cytochrome P450 enzymes, stratification of QT-DDIs is extremely complex and larger studies need to be conducted for each QT-DDI separately [
44,
45].
The performance characteristics of the model were not perfect. Also after performing a post-hoc analysis, the discrimination ability of the model remained limited. This can be explained by the discrepancies between the development and validation cohort. First, the validation dataset included patients from all departments including ICU patients, whereas the development dataset only included patients from medical wards. Unfortunately, we could not exclude these patients in the validation dataset, because it was unknown to which department patients had been admitted. Therefore, we decided to exclude patients using propofol in order to exclude perioperative and ICU patients as much as possible. Also, patients with ICDs or ECG abnormalities were not excluded in the validation cohort because these data could not be extracted. Therefore, we excluded ECGs with deviant heart rates and QTc-intervals. We did correct the QT-interval for wide QRS-complexes to limit ECG exclusions. Second, the QTc-intervals of the development cohort were manually measured, while the QTc-intervals of the validation cohort were automatically calculated by the MUSE Cardiology Information System. But most importantly, the retrospective design of the external validation where only patients in whom an ECG was recorded during use of the QTc-prolonging drugs were included, may have led to selection bias. ECGs are more likely to be recorded in vulnerable patients. According to the high prevalence of comorbidities in the validation cohort, this was probably the case. But even in high risk populations, QTc-prolongation is not always present resulting in false positives. Also, ECGs are more likely to be recorded in patients with underlying cardiac diseases or with suspected QTc-prolongation even if they only have a few risk factors, resulting in false negatives. Our preliminary results must therefore be confirmed in large studies where this selections bias is not present. The usability of the tool must be evaluated in a clinical setting. For future perspectives, this tool must be further studied to assess its effect when it is integrated in an electronic decision support system before implementation can be recommended. A clinical decision support system is extensively used by pharmacists, as it is part of their job to read DDI alerts. Ideally, the system will automatically calculate a risk score for the individual patient and only generate alerts in high-risk patients resulting in more specific alerts. Such a study should be performed in large patient groups with clinically relevant endpoints.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.