Background
Depression is a severe affective mental disorder that is accompanied by a lack of pleasure, and the impairment of cognition, behavior and autonomic nerve function, which causes dysfunction in various spheres of individual and social life, severely limits psychosocial functioning, and diminishes quality of life [
1,
2]. Being one of the most widespread, pervasive, and troublesome illnesses in the world [
3‐
5], depression can affect individuals of any age. By 2020, depression is anticipated to overtake heart disease as the second-leading cause of disability or early death, according to estimates from the World Health Organization (WHO) [
6]. As a common and disabling mental disorder [
7], it is a serious global public health issue that not only results in personal misery for those affected but also places a large economic burden on both the patients and the entire society [
8,
9]. When it comes to medicinal therapy for depressive disorders, the American Psychiatric Association recommends selective serotonin reuptake inhibitors (SSRI, such as sertraline) and serotonin–norepinephrine reuptake inhibitors (SNRI, such as duloxetine), as well as noradrenergic and specific serotonergic antidepressants (NaSSA, such as mirtazapine) [
10,
11]. Regrettably, despite obtaining evidence-based antidepressant medication, up to 70% of people are going to continue to experience troublesome symptoms [
12,
13].
According to the Canadian Network for Mood and Anxiety Treatments (CANMAT) guidelines and American Psychiatric Association Practice guidelines, atypical antipsychotics (AA), specifically the use of quetiapine has been reported as an effective augmentation strategy to antidepressants. Quetiapine is an atypical antipsychotic agent, which was first introduced in the pharmaceutical market in 1997 [
14]. In 2010, the European Medicine Agency (EMA) approved the extended-release formulation of the drug, quetiapine XR, as an add-on to antidepressants when monotherapy gives suboptimal response [
15]. Studies have shown that quetiapine (mean dose, 156.74 ± 97.6 mg/day) showed significant benefits for both response and remission rates compared to placebo [
16,
17]. Despite its high effectiveness, its optimal use is limited by widely variant individual factors, including height, weight, age, medical history, and the CYP3A4 and CYP2D6 enzymes and so on [
18]. Before achieving the quetiapine maintenance dose, these influencing factors make it challenging to reach the narrow therapeutic window, which is monitored by the therapeutic drug monitoring from AGNP and a sub- or supra-therapeutic recommended therapeutic reference range (200–750 ng/ml). This may render treatment ineffective or increase the risk of sedation, hypotension, dry mouth, constipation, and tachycardia. Therefore, it is critical to help clinicians select the appropriate quetiapine dose and individualized quetiapine treatment using prediction models.
Recently, there has been a trend toward using machine learning and deep learning methods to create customized medications based on research from real-world situations [
19]. With the help of large-scale complex algorithms and datasets, machine learning and deep learning algorithms, a branch of artificial intelligence, are able to predict clinical outcomes with high accuracy [
20,
21]. When predicting from a variety of variables, they can assess data-driven estimation and derive nonlinear variable linkages [
20,
21]. Several studies have utilized machine learning and deep learning approaches to improve the model depiction of the complex link between individual characteristics and drug dose, such as a vancomycin treatment prediction system using Extreme Gradient Boosting (XGBoost) [
22], and a brand-new warfarin maintaining dose prediction system using Light Gradient Boosting Machine (LightGBM) [
23].
Herein, our goal was to build a prediction model of quetiapine adjusted dose in a stationary state using algorithms based on machine learning and deep learning to support clinical prescription decisions. We did this by maximizing the use of real-world data to find significant influencing variables for quetiapine dose.
Discussion
One of the most popular and efficient ways to treat depression in a therapeutic context is with antidepressants, which also has the ability to successfully slow the onset of disease in depressed patients. However, research indicates that only about half of major depressive disorder (MDD) patients receive antidepressants that work well for them, and only about a third of them experience remission [
27]. The use of AAs as first-line medicines, notably quetiapine, is recommended by various current pharmacological augmentation guidelines for treating depression [
17,
28,
29]. The right quetiapine dose and personalized quetiapine treatment are frequently challenging for clinicians.
To better estimate quetiapine dose during depression treatment and to find valid and accurate predictors, we compared the prediction abilities of quetiapine dose by applying nine machine learning and deep learning techniques for patients with depression. Ultimately, the XGBoost algorithm with the best performance (accuracy = 0.69) among nine models was selected to build the prediction model. Afterward, it can be observed that a number of 43 instances of quetiapine dose were properly predicted in the testing cohort. The overall accuracy of the model was 0.69. The moderate accuracy demonstrates that the effect of accurately predicting quetiapine dose is acceptable, and our findings may offer clinicians recommendations for prompt drug regimen adjustments. In addition, we performed dose subgroup analyses to show individual predictive performance across dose levels and to help refine model performance with continued recruitment of data for a given range of daily doses.
Calculations of the area under the concentration–time curve (AUC), for instance, provide the basis of classic pharmacokinetic studies. However, if the data are insufficient or cannot support a pharmacokinetic modeling technique, the model is erroneous [
30]. Recently, it has been noticed that there is growing interest in novel statistical techniques, such as population pharmacokinetic (popPK) analysis. Nonlinear mixed-effects modeling (NONMEM) is the most popular method for this type of pharmacokinetic data analysis [
31,
32]. Nevertheless, the PPK model is relatively inflexible to apply because of the explicit mathematical model used, and adding or removing a parameter may be challenging [
33]. Machine learning, in contrast, is renowned for its self-organizational and learning skills, which let computers learn from “experience” without being explicitly taught [
34,
35]. It is a form of artificial intelligence that enables systems to examine a wide range of data gathered from electronic health records (EHRs) and automatically learn from them using cutting-edge statistical and probabilistic techniques to make more precise predictions by building clever and efficient predictive models [
36]. Recent years have seen a significant increase in study interest in the use of machine learning for clinical drug therapies, which leads to an increasingly significant impact on the development of personalized dosing, particularly in the choice of drug dose [
37]. A few studies on the use of machine learning to forecast drug doses or blood concentrations have been reported [
38‐
43].
In this study, we innovatively used machine learning and deep learning techniques to predict quetiapine dose based on real-world data. Machine learning models can be updated by automatically extracting EHR data and continuously monitoring physiological data, and are effective approaches to modeling real-world data. The commonly used PPK models have some limitations, such as difficulty in modeling, less consideration of influencing factors, and low accuracy. Herein, multi-level data mining was conducted by machine learning to screen out a variety of real-world influencing factors, to construct a more practical and accurate dose prediction model. Therefore, the combination of machine learning and dose prediction can help to improve the level of precision medicine in clinical.
We considered multiple algorithms for model establishment. DT is simple and easy to understand, but there is a risk of overfitting. RF uses bagging sampling, random attribute selection and model ensemble to address excessive risk decision tree learning. On the basis of RF, GBDT combined with Boosting establishes the connections between trees, making the forest an ordered collective decision-making system. XGBoost goes a step further than GBDT by adding regular terms to the objective function at each iteration to reduce the risk of overfitting, and it can integrate multiple decision trees to achieve the goal of regression or classification [
44]. For models such as ANN and XGBoost, they perform quite well on large-scale datasets. However, good prediction results can also be obtained on small data sets by adjusting hyperparameters to avoid overfitting. Each algorithm has its advantages and disadvantages, the performance of different algorithms depends on the characteristics of the dataset, and the final selection of the algorithm is based on the computational results. Herein, we used grid search combined with tenfold cross validation to find the optimal hyperparameters and avoid overfitting to obtain the optimal model.
The significant predictor for predicting quetiapine dose is the quetiapine concentration. Several studies on psychotic disorders have identified that dose affects quetiapine concentration. According to a review, quetiapine had linear pharmacokinetics in the studied dose range, and had predictable pharmacokinetics [
45]. Albantakis et al. have also quantified the relationship between daily dose and serum concentration in children and adolescents with psychotic and mood disorders. Between the daily dose and quetiapine serum levels (from trough samples) in the entire sample, they discovered a statistically significant, positive, but flimsy linear connection [
46]. Among the crucial parameters we chose for our study's prediction model, the concentration was the most prominent influencing variable, and it was positively associated with quetiapine dose, which was in line with earlier research.
The effect of age on the metabolism of second-generation antipsychotics has been described in a few prior investigations. One study revealed that dose-adjusted concentrations of quetiapine increased by an average of 13% per decade from the age of 20 [
47], while another found that the average concentrations were 67% higher in patients over the age of 70 compared to those between the ages of 18 and 69 [
48]. Another study found that patients aged 65 and above had 50% higher plasma concentrations than younger patients [
49]. For children and adolescents (10–17 years of age), at steady state, the pharmacokinetics of the parent compound were similar to adults. However, when adjusted for dose and weight, AUC and Cmax of the parent compound were 41% and 39% lower, respectively, in children and adolescents than in adults [
50,
51]. In our study, patients older than 60 years were excluded because of the small number of senior patients that model can only learn little information. The ability of the elderly to metabolize and excrete drugs may be reduced, which may lead to the accumulation of drugs in the body, and liver and kidney function may also be affected. As a result, older people tend to require smaller doses of drugs. In this study, age is one of the most important feature in the final prediction model. In the following study, we will include more patients older than 60 years in the model to verify its generalizability.
In addition, some previous studies have indicated that low MCHC increases the likelihood of developing pathological disorders, such as poor functional status, dementia, and cognitive decline as well as morbidity and death [
50‐
53]. Poor functional status, such as decreased ability to carry oxygen, may lead to changes in the pharmacokinetics of quetiapine and thus affect the dose of quetiapine. Meanwhile, because it is a measure determined from the haemoglobin concentration (HGB) divided by mean cellular volume (MCV) and red blood cell count (RBC), the MCHC is a good indicator to detect anaemia [
54]. Depending on the demographic data investigated, anemia, a condition marked by a deficiency in hemoglobin in the blood, affects an estimated percentage from 2.9% to 60.1% of older persons [
55]. Many illnesses, including malnutrition, obesity, cancer, chronic renal disease, are linked to anemic people, which may lead to changes in the pharmacokinetics of quetiapine and thus affect its dose.
Furthermore, hepatic metabolism accounts for the majority of quetiapine elimination, and less than 1% of the amount taken orally after a single administration was excreted unaltered, showing quetiapine is rapidly metabolized [
56,
57]. According to studies, people with liver disease (n = 8) had a 30% lower mean oral clearance of quetiapine than patients with normal liver function. Two of 8 patients with hepatic impairment experienced a threefold increase in AUC and Cmax compared with healthy patients [
56,
57]. TBA is closely related to liver function and abnormally high value suggests poor liver health. Abnormal TBA levels indicate that patients may have impaired liver function, which may inhibit metabolism of quetiapine in the liver, resulting in high quetiapine concentration and dose adjustment may be needed. In one word, quetiapine TDM value, age, MCHC, and TBA, show important associations with quetiapine dose, which could be used as the predictors in the individualized medication model of quetiapine, to help clinicians choose the reasonable regimen.
In different dose groups, according to Additional file
1: Figure S1, the blood concentration points of some patients with a dose of 200 mg are extreme outliers, and there is a crossover with the upper quartile concentration points of patients with a dose of 400 mg. Also, there is a crossover between the upper quartile concentration points of patients with a dose of 200 mg and the lower quartile concentration points of patients with a dose of 300 mg. All the situations of crossover may affect the clinician’s regimen choice and the prediction outcome in 200 mg group. Therefore, the AUROC for 200 mg group is lower than other dose groups.
Our model has some notable flaws. First, due to the availability of data, such as extremely uneven distribution, lots of missing values and so on, some variables were excluded. A future goal is to improve the model when a great deal of samples may be used to thoroughly study the factors. Second, our model has not been sufficiently tested on additional data sets. By using the model on a larger pooled data set, future studies could delve deeper into these problems. The identification of more potent predictors and the improvement of prediction accuracy are likely to result from the input of additional data. Third, due to the constraints of the test conditions, several pertinent patient characteristics (such as CYP450 polymorphisms) were excluded. Last, in this study, some underlying confounding factors were not analyzed, such as the using duration of quetiapine, prior use of antipsychotics, mood stabilizers and antidepressants before admission, drug combination of benzodiazepines, anxiolytics, and lithium, and complex clinical situations including severity of illness and multiple complications [
58]. There is a drawback of real-world study that there exist some unknown confounders from real clinical settings. In future study, we expect to apply propensity score matching and stratified analysis for reducing confounding bias.
According to our knowledge, this research is the initial to use XGBoost algorithm for estimating the dose of quetiapine for patients with depression. Our study could identify important influencing variables for quetiapine dose by maximizing the use of real-world data to support quetiapine dose adjustments for each patient. In clinical applications, we expect to develop a web tool for drug dose calculation that can automatically generate recommended quetiapine doses by entering the values of key variables (such as quetiapine TDM value, age, MCHC, and TBA) based on electronic medical records, blood tests and TDM, providing clinical decision support to improve therapeutic response and reduce patient’s burden.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.