Producing personalized statin treatment plans to optimize clinical outcomes using big data and machine learning

Abstract

Almost half of Americans 65 years of age and older take statins, which are highly effective in lowering low-density lipoprotein cholesterol, preventing atherosclerotic cardiovascular disease (ASCVD), and reducing all-cause mortality. Unfortunately, ~50% of patients prescribed statins do not obtain these critical benefits because they discontinue use within one year of treatment initiation. Therefore, statin discontinuation has been identified as a major public health concern due to the increased morbidity, mortality, and healthcare costs associated with ASCVD. In clinical practice, statin-associated symptoms (SAS) often result in dose reduction or discontinuation of these life-saving medications. Currently, physician decision-making in statin prescribing typically relies on only a few patient data elements. Physicians then employ reactive strategies to manage SAS concerns after they manifest (e.g., offering an alternative statin treatment plan or a statin holiday). A preferred approach would be a proactive strategy to identify the optimal treatment plan (statin agent + dosage) to prevent/minimize SAS and statin discontinuation risks for a particular individual prior to initiating treatment. Given that using a single patient’s data to identify the optimal statin regimen is inadequate to ensure that the harms of statin use are minimized, alternative tactics must be used to address this problem. In this proof-of-concept study, we explore the use of a machine-learning personalized statin treatment plan (PSTP) platform to assess the numerous statin treatment plans available and identify the optimal treatment plan to prevent/minimize harms (SAS and statin discontinuation) for an individual. Our study leveraged de-identified administrative insurance claims data from the OptumLabs^® Data Warehouse, which includes medical and pharmacy claims, laboratory results, and enrollment records for more than 130 million commercial and Medicare Advantage (MA) enrollees, to successfully develop the PSTP platform. In this study, we found three results: (1) the PSTP platform recommends statin prescription with significantly lower risks of SAS and discontinuation compared with standard-practice, (2) because machine learning can consider many more dimensions of data, the performance of the proactive prescription strategy with machine-learning support is better, especially the artificial neural network approach, and (3) we demonstrate a method of incorporating optimization constraints for individualized patient-centered medicine and shared decision making. However, more research into its clinical use is needed. These promising results show the feasibility of using machine learning and big data approaches to produce personalized healthcare treatment plans and support the precision-health agenda.

1. Introduction

Statins are highly effective drugs to reduce low-density lipoprotein cholesterol and the risk of nonfatal and fatal atherosclerotic cardiovascular disease (ASCVD) events. Current U.S. guidelines recommend moderate or high-intensity statin therapy for patients most likely to benefit; that is, those with ASCVD, as well as those with low-density lipoprotein-C ≥ 190 mg/dL, diabetes, 40–75 years of age, or ≥ 7.5% 10-year ASCVD risk [1,2]. During the 2007–2010 time period, approximately 47% of Americans 65 years of age and older were prescribed statins [3]. Moreover, the rate of statin use has increased about 7-fold since 1988–1994, largely due to the acceptance of the effectiveness of statins to reduce ASCVD [4].

Unfortunately, ~50% of patients prescribed statins discontinue their use within one year of treatment initiation [5]. Such discontinuation leads to the loss of the ASCVD-prevention benefit for a vast number of Americans; accordingly, statin discontinuation has been identified as a significant public health concern due to the increased morbidity, mortality, and healthcare costs associated with ASCVD [6–11]. Therefore, it is not surprising that a study among veterans with ASCVD concluded low statin therapy adherence was associated with a higher risk of death [12].

Although statins are, in general, well-tolerated, this class of drugs has been found to be associated with a slightly increased risk of diabetes, hemorrhagic stroke, and severe myopathy in clinical trials [13,14]. Observational studies have reported an association between statins and increased rates of skeletal muscle [15], central nervous system, liver, renal, and sleep dysfunction, as well as other possible statin-associated symptoms (SAS) [16]. In a large Internet survey (USAGE), 62% of former statin users reported that they stopped statin use because of SAS [5,16,17]. Notably, SAS is more commonly reported in clinical practice (~20% of statin users) than in randomized controlled trials (~5% of statin users), which could be due to more careful patient selection criteria inherent in the latter [18,19]. Similarly, in one study among ASCVD patients, approximately 55% of patients (former statin users) cited perceived side effects as a reason to stop their statin therapy [15]. As a result, SAS has been cited as the most common cause of statin dose reduction or discontinuation [6,16].

In our preliminary observational study using the OptumLabs^® Data Warehouse (OLDW), we found that patients with SAS were more than three times more likely to discontinue statin therapy than those who did not report SAS. In current clinical practice, “reactive” strategies are typically employed to manage statin intolerance due to SAS. These include changing the statin treatment plan (e.g., reducing dosage or changing medication), using a statin holiday (short-term discontinuation), or trying alternative (non-statin) cholesterol-lowering therapy [19,20]. These current strategies can be problematic because switching to a lower-intensity statin treatment plan might reduce SAS risk and reduce statin use’s clinical benefits [19]. Switching to non-statin therapy is also problematic, given that many non-statins have worse adverse events and poorer ASCVD risk reduction than statins [19,21]. These traditional approaches to manage SAS after they manifest can involve trial and error or “best guess” treatment plan changes in clinical settings, leading patients to discontinue statin treatment of blood cholesterol levels altogether.

It is not surprising that the current reactive approach to manage SAS is problematic, given that it is only initiated after problems have emerged. In addition, the current statin prescribing approach that relies on only a few patient data elements (such as comorbidity or age) or uses only that single patient’s data to identify the optimal statin regimen and treatment plan is inadequate to ensure that the harms of SAS are minimized. This report proposes and explores a model to predict the personalized statin treatment plan (PSTP) for a particular individual that prevents or minimizes their SAS. Specifically, we first train the model using a variety of variables from a large number of patients (“big data”) and then use the trained model to predict the optimal PSTP for individuals based on the individual’s specific data and characteristics. In other words, this model leverages big data to produce an individualized, proactive prescribing strategy based on the individual’s data obtained before the initial statin prescription is made.

In addition, we will also explore two features that further help improve the PSTP’s practical value. First, in addition to minimizing SAS, the PSTP also aims to mitigate the risk of statin therapy discontinuation. Therefore, we will discuss our approach to finding a multi-objective optimization solution and provide initial evidence that the PSTP shows significantly lower risks in both SAS and statin therapy discontinuation compared to the treatment plan received initially by patients. Second, our model explicitly considers the individual patient’s preferences to support patient-centered medicine and shared decision-making,

2. Materials and methods

There are 4 core components: (1) cohort extraction, (2) outcome definition, (3) prediction, and (4) optimization. To show the method’s feasibility, (1) focuses on extracting data for secondary prevention and have not taken statin before, (2) discusses risks of SAS and therapy discontinuation, and one aims to identify treatment plan with minimum risks, (3) discusses the development of prediction models for SAS and discontinuation, and (4) discusses three types of optimization to identify PSTP - single optimization, optimize both SAS and discontinuation, and patient-centered optimization.

3. Cohort extraction

3.1. Data source

We conducted our study using commercial and Medicare Advantage insurance claims data in the OLDW, which includes HIPAA compliant de-identified medical and pharmacy claims, laboratory results, and enrollment records for more than 200 million commercial and Medicare Advantage (MA) enrollees from 1993 to the present [22]. The database contains longitudinal health information on enrollees and patients, representing a mixture of ages, ethnicities and geographical regions across the United States. In addition, various healthcare economic [24]; comorbidity [25], treatment approaches [26], patterns of drug use [27], health outcomes [28], drug interaction [29], data observations for statin prescription and toxic reaction [30] types of studies have used OLDW data and have been published in peer-reviewed journals. Among these studies, our preliminary studies [29,30] provide many observations in the same data cohort used in this study, which further support the need for PSTP and proactive strategy development. This data source provides the substantial capacity to do viable subgroup analysis and the ability to discern patterns in support of better clinical decision-making. The Institutional Review Board (IRB) at the University of Minnesota reviewed and approved the study (ID: STUDY00004125).

3.2. Years of data selection

The database contains multi-year data back to the year 1993. Selecting reasonable and manageable years of data is crucial to the practicability of our study. All of the data from 2010 to 2015 were included to enhance the generalizability of our model to current medical practice. Medicare Part D went into effect in 2006. Its implementation affected pharmaceutical coverage prices and increased the utilization of prescription medications, with changes in coverage from insurance plan to Medicare Part D. However, the seven statins (atorvastatin, lovastatin, fluvastatin, pravastatin, pitavastatin, simvastatin, and rosuvastatin) of interest to our study were all in use by 2010. Pitavastatin, the newest statin in the United States, received FDA approval in 2009 and was brought to the market in 2010. Therefore, filled prescriptions of all available statins can be obtained after 2010. In addition, the Affordable Care Act was enacted in 2010. The Affordable Care Act changed healthcare in many respects, such as increasing insurance coverage, changing insurance standards, and affecting insurance premiums and healthcare costs. Together, these reasons made it more likely that our findings could be generalized to patient care in the present.

3.3. Inclusion and exclusion criteria

Statins are the preferred medication for ASCVD patients to manage lipids for secondary prevention. To support our research, we selected a patient cohort (Fig. 1) from the OLDW based on the following criteria:

Fig. 1. Selection of the study population from the OLDW.

CE: continuous medical and pharmacy enrollment.

• Initially, patients with ASCVD diagnoses from 2010 to 2015 were selected from the entire OLDW claims database. However, patients with radiation-induced coronary artery disease or heart transplant-related coronary artery disease were excluded, given that the coronary artery disease mechanism and treatment strategies are different for these diseases than those from other ASCVD. The first ASCVD diagnosis date, occurring within the 2010–2015 data subset, was defined as the ASCVD index date. Diagnosis and procedure codes related to Myocardial Infarction; Acute and Chronic forms of Ischemic Heart Disease; Cerebral Infractions; Transient Ischemic Attacks were used to define ASCVD patients.

• The data was further filtered to include ASCVD patients of 40 + years and older. The rationale for this selection criteria was based on studies that claimed the prevalence of ASCVD in the age group 40 years and above increases substantially from 10% to 40% based on a national survey from 2009 to 2012 [31]. Another study using NHANES data concluded that in the absence of risk factors, women aged 50 years and younger and men aged 40 years and younger had a low risk of ASCVD prevalence [32].

• Patients with less than one year of continuous insurance enrollment before the ASCVD index date were excluded to reduce the risk of inaccurate index date identification due to left censoring. This criterion ensured that the patients were less likely to have a prior history of ASCVD or statin use. For patients with an ASCVD index date during 2010, further exclusion criteria were applied to ensure no other ASCVD diagnoses were made in the previous year before their index date. Previous studies have also used a one-year lookback period to accurately capture new statin users and ASCVD patients [23]. Patients who had follow-up data for one year after the statin index date were also included to capture adverse events that are likely to occur early (3–6 months) in the therapeutic course. Because of these criteria, there was no need to censor data.

• Patients who filled an index/first statin prescription within 30 days after the ASCVD index date were also included. Statin index date was defined as the date of the first statin prescription after the ASCVD index date. The following statin drugs were included in our study: atorvastatin, cerivastatin, fluvastatin, lovastatin, pitavastatin, pravastatin, rosuvastatin, and simvastatin.

• Patients with a statin index date before 09/01/2014 were included. At the time of our analysis, this restriction on the statin index date ensured that we would have access to patients’ medical claims occurring up to one year after their statin index date. Using this cut-off time, the statin index dates of the included patients were between 01/01/2010 and 08/31/2014. Additionally, patients with one year of continuous medical and pharmacy enrollment before the ASCVD index date and patients with one year and one month of continuous medical and pharmacy enrollment after statin index date were included. One year of continuous enrollment allowed us to ascertain that a patient had no prior history of statin use before the CVD index date. After a one-year follow-up, an extra month of continuous enrollment was needed to differentiate continuous and discontinued patients (as described in the treatment plan and adherence behavior section below). The extra month of statin claims was also used to categorize patients based on their statin continuation status.

• Statin users who were using combination statin drugs were excluded from this study. Similarly, duplicate records for patients were also removed.

• Finally, patients who did not maintain the same treatment plan were excluded. This step reduced the bias in the prediction model due to the treatment plan transition.

4. Outcome definition

4.1. Treatment plan and adherence behavior

Previous studies have shown that the risk of adverse events in statin use is associated with specific statin agents [19,33,34] and statin dosage [35–37]. Empiric observation yields our cohort’s 27 possible statin treatment plans (statin agent + dosage) (Table 1). These treatment plans utilized seven different statins and three dosing intensities (low, moderate, and high). In other words, these 27 different statin treatment plans had been prescribed in clinical settings and reimbursed by insurance.

Table 1.

Statin Dosage Intensity Groupings.

Statin	Intensity
	Low	Moderate	High
Atorvastatin	–	10 mg, 20 mg	40 mg, 80 mg
Simvastatin	5 mg, 10 mg	20 mg, 40 mg	80 mg
Pravastatin	10 mg, 20 mg	40 mg, 80 mg	–
Rosuvastatin	–	5 mg, 10 mg	20 mg, 40 mg
Lovastatin	10 mg, 20 mg	40 mg	60 mg
Pitavastatin	1 mg	2 mg, 4 mg	–
Fluvastatin	20 mg, 40 mg	80 mg	–

Patients’ medication adherence behaviors were classified into two groups for the prediction modeling: continuous and discontinued. Patients who continuously used a statin for at least one year without any gap longer than 30 days were categorized into the continuous group. Otherwise, patients were categorized in the discontinued group. The gap was calculated as follows:

$$\mathrm{G}\mathrm{a}\mathrm{p}\mathrm{d}\mathrm{a}\mathrm{y}\mathrm{s}=2\mathrm{n}\mathrm{d}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{n}\mathrm{f}\mathrm{i}\mathrm{l}\mathrm{l}\mathrm{d}\mathrm{a}\mathrm{t}\mathrm{e}-(1\mathrm{s}\mathrm{t}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{n}\mathrm{f}\mathrm{i}\mathrm{l}\mathrm{l}\mathrm{d}\mathrm{a}\mathrm{t}\mathrm{e}+1\mathrm{s}\mathrm{t}{\mathrm{f}\mathrm{i}\mathrm{l}\mathrm{l}}^{\prime }\mathrm{s}\mathrm{d}\mathrm{a}\mathrm{y}\mathrm{s}\mathrm{o}\mathrm{f}\mathrm{s}\mathrm{u}\mathrm{p}\mathrm{p}\mathrm{l}\mathrm{y})$$

4.2. Statin-associated symptoms (SAS)

This proof-of-concept study defined SAS using International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) codes. These SAS are rhabdomyolysis (728.88), myopathies (359.4, 359.8, 359.9, 728.8, 728.87, 728.89, 728.9, 729.1, 791.3, 729.89), renal events (584.XX), liver events (570, 573.X, 790.5), and poisoning events (972.2, E942.2, E980.4). These are more severe SAS. On the other hand, minor SAS (muscle pain) is more likely to appear in clinical notes than claims. Because these minor SAS are often non-billable and not a primary cause of a visit to a physician, while these minor SAS might have been documented in clinical notes, clinical notes were not available in this claims-based study. Because our model was intended to predict more severe SAS (that is, those with an associated ICD-9-CM code) and recommend the PSTP that prevents or minimizes risks of the severe SAS, it was not impacted by the lack of access this documentation.

4.3. Prediction

4.3.1. Data preprocessing

We log-transformed skewed variables and had a wide distributional range to obtain a normal distribution [38]. Some continuous variables that are measured in different scales were not directly comparable and did not have equal contributions to prediction. For instance, age can have a larger value than statin dosage. To reduce the chance that these variables with larger values dominate the prediction model, we further used feature scaling to rescale all continuous values by subtracting the mean and dividing by the standard deviation [38].

4.3.2. Prediction model for SAS and statin treatment discontinuation

Our research cohort used in this study includes 72 variables for 38,214 patients. Among these variables, 70 predictors (listed out in the supplemental materials) included basic demographics (e.g., age and gender), treatment (e.g., drug type and dosage), healthcare claims, and comorbidity (see supplement 1 for specific variables). Two target binary variables were used: SAS (0 – no SAS; 1 – any observed SAS); and observed discontinuation (0 – continue; 1 – discontinue statin treatment). Both SAS and discontinuation outcomes are close to some degree. Discontinuation is a critical issue in statin therapy, and SAS is the most important reason. On the other hand, some patients continue statin treatment even when they have SAS. As a result, there is partial overlap between these two outcomes, and either one is important. We note that the prediction only uses variables that can be observed before the prescription to predict future SAS and discontinuation. Some variables may have different values at different time points, such as age. In this case, we use the age value at the time of prescription to make the prediction.

We used an artificial neural network model to extract predictive signals for risks of SAS and statin treatment discontinuation from the data and used the model to estimate and predict these risks. We applied Neural Network Pattern Recognition toolkit to develop the artificial neural network model in MATLAB computational framework. Due to limited computational resources, we were looking for a neural network architecture that is necessary and sufficient to capture predictive signals and simple enough to be efficient in terms of computational expenses associated with its training and practical use. Therefore, the final chosen architecture was a feed-forward neural network [38] with sigmoid hidden and softmax output. The network was trained with seed conjugate gradient backpropagation with cross-entropy as the loss function. The number of the epoch is 22, which has the highest performance. We empirically tested the predictive performance of different numbers of hidden layers and found that one hidden layer was sufficient. In addition, we empirically tested and explored the number of neurons in the hidden layer between 70 (number of input nodes; each corresponds to a predictor) and 2 (number of output nodes). Finally, we set the number of neurons to 10 in the hidden layer as the architecture balances computational costs and performance improvement (Fig. 2).

Fig. 2.

Artificial Neural Network architecture for estimating statin treatment risks.

We note that the neural network method learns prescription patterns from those 70 variables. Then we use early stopping [38] based on the cross-entropy function as the regularization method to control overfitting. In other words, we do not pre-select significant variables to develop the model. Alternatively, we let the model determine sufficient predictive information obtained from these variables with the regularization method. Additional regularization methods (such as dropout, weight decay, and L1 or L2 norm regularization) were only needed if we noticed a significant difference in model performance between the testing and training data. That indicates the model has an “overfit” to spurious correlations present in the data. As a result, we noticed no significant difference in Precision-Recall AUC (PR-AUC) metrics over testing and training data to indicate any overfitting. Therefore, we did not incorporate any additional regularization methods apart from early stopping.

This optimized configuration of a neural network allowed us to scale our computational methods and to process more variables and more samples while at the same time avoiding drastic increases in computational expenses. We constructed separate neural networks to predict SAS risk and treatment discontinuation risk.

We randomly divided the extracted data into three parts: 70% training set, 15% validation set, and 15% testing set. The training set was used to adjust or fit the network to the input data. The validation set was used to tune parameters, such as the number of hidden units, and the testing set was used to provide an independent measure for network performance on previously “unseen” data.

4.3.3. Using prediction and optimization to identify the PSTP

Identifying the optimal PSTP for an individual occurs in two steps. First, for a given patient, make a prediction for the value of each outcome variable for each possible treatment. Second, apply an optimization method to the predicted outcomes to identify the optimal PSTP.

In the present study, we are using the trained neural network to make the predictions (i.e., 27 values for each outcome variable are predicted by the trained neural network based on a patient’s set of predictor values and cycling through the 27 possible treatment options as input). Making the predictions provides a basis for choosing one out of the 27 possible statin treatment plans for a patient. With the full schedule of predictions in hand, plan selection becomes a matter of translating patient and physician preferences into the mathematical language of an appropriate optimization method.

If there is only one outcome variable of interest, identifying the optimal plan is straightforward; however, it becomes a multi-objective optimization problem with many potential solutions when there are two or more outcome variables of interest. In either case, patient and physician goals can be translated into optimization constraints to identify the ideal PSTP. This study demonstrates both single optimization (using SAS or discontinuation risk as to the single outcome of interest) and multi-objective optimization (using both SAS risk and discontinuation risk as to the outcomes of interest). The concept of integrating optimization and prediction to identify the optimal health decision has been discussed in our previous work [39,40].

Once the above prediction model has been trained and developed, the following Eqs. (1) to (3) were further used to optimize and identify the PSTP. Specifically, these objective functions aim to identify: (1) PSTP with either optimized SAS or discontinuation, (2) PSTP with both optimized SAS and discontinuation, and (3) not only PSTP with both optimized SAS and discontinuation but also satisfying personal preference constraint.

4.4. Optimization

4.4.1. Using single-objective optimization to identify the PSTP

To identify the optimal PSTP that minimizes the risk of SAS, we first separated each patient’s record ($x$) into unchangeable ($x_1$) and changeable ($x_2$) variables. In this case, $x_2$ is the treatment plan, which includes 27 possible options. We aimed to find the option with the minimal risk of SAS (or maximal confidence of a low SAS risk), so we sought to identify the optimal solution ($x_2^{\mathrm{*}}$) with the minimal predicted SAS risk. We described the optimization formulation as follows:

$$\begin{array}{c}\underset{x_{2i}}{\mathrm{m}\mathrm{i}\mathrm{n}}\left(f_{SAS}\left(x_1+x_{2i}\right)\right)\\ \mathit{st}.i=1,\cdots ,\left|x_2\right|,\end{array}$$ (1)

where $f_{SAS}$ is the function (or the above artificial neural network model) to predict SAS. Among treatment plan options, we aimed to identify the option $x_2^{\mathrm{*}}$ (PSTP) with the minimal $f_{SAS}$.

4.4.2. Using multi-objective optimization to identify the PSTP

The equation cited above only identifies the optimal treatment plan based on minimal SAS. However, the PSTP can be produced under different types of clinical scenarios. For example, to add the risk of treatment discontinuation in the PSTP selection process, we added a new optimization objective and explored the minimal distance for the two risks as a multi-objective optimization [41,42]. The optimization problem is:

$$\begin{array}{c}\underset{x_{2i}}{\mathrm{m}\mathrm{i}\mathrm{n}}\left(\sqrt{f_{SAS}{\left(x_1+x_{2i}\right)}^2+f_{Disc}{\left(x_1+x_{2i}\right)}^2}\right)\\ \mathit{st}.i=1,\cdots ,\left|x_2\right|\end{array}$$ (2)

Here, we determined the PSTP is the treatment plan with the optimal balance of SAS ($f_{SAS}$) and discontinuation ($f_{\mathit{Disc}}$). Because the solution space is very small (only 27 treatment plan options), an exhaustive search is sufficient to conduct this optimization task. We note that Eqs. (2) and (3) aim to identify the treatment plan with the lowest risks of SAS and discontinuation, but not all cases can simultaneously have the lowest risks for both SAS and discontinuation. Alternatively, we use Euclidian distance to balance both, which is the minimum Euclidean of both. Minimizing the Euclidean distance measure is equivalent to the patient expressing no preference between minimizing SAS risk and minimizing discontinuation risk.

4.4.3. Patient-centered optimization of the PSTP

The previous two equations evaluated all possible treatment plans for an individual. However, for practical reasons, an individual may have a particular reason or preference to avoid certain types of treatment plans (e.g., cost, clinical, or other personal preferences). To address this clinical scenario, we added constraint(s) to the above multi-objective optimization. The optimization problem is:

$$\begin{array}{c}\underset{x_{2i}}{\mathrm{m}\mathrm{i}\mathrm{n}}\left(\sqrt{f_{SAS}{\left(x_1+x_{2i}\right)}^2+f_{Disc}{\left(x_1+x_{2i}\right)}^2}\right)\\ st.i\in personalpreference\end{array}$$ (3)

The constraint is specific to an individual who is limited to a particular type of statin treatment plan. We will present an example of patient-centered PSTP optimization in the Results section.

We note that Eqs. (2) and (3) aim to identify the treatment plan with the lowest risks of SAS and discontinuation (represented by $f_{SAS}$ and $f_{\mathit{Disc}}$). To do this, we use the Euclidian distance function to identify the treatment plan with the lowest risks (represented by the minimum Euclidean distance).

Table 2 summarizes the abovementioned predict and optimization methods/module. For example, one needs prediction data and physician-patient preference and puts them in Eq. (3) to produce Fig. 4.

Table 2.

Summary of integrating prediction and optimization modules to identify PSTP.

	Prediction module	Optimization module
Purpose	Generate risk estimation treatment-plan data to facilitate the following optimization task	Identify PSTP
Needed data to develop the module	OptumLabs Data Warehouse (Total n = 38,214)	(1) Prediction data for 27 treatment plans (27 possible solutions for each individual) and (2) Physician-patient preference (served as hard constraints used in Eq. (3) to reduce the number of solutions further)
Module example in this manuscript	Artificial neural networks (ANN) for two sets of binary classification: SAS and discontinuation	(1) Eq. (1) (either SAS or discontinuation is optimized) or (2) Eq. (2) (both SAS and discontinuation are optimized) or (3) Eq. (3) (satisfying physician-patient preference constraint and both SAS and discontinuation are optimized)

Fig. 4.

Example of a personalized patient-centered statin treatment planning decision plot.

5. Results

We will first discuss the rationale for using machine learning to develop PSTPs, and then present a comparative analysis of original vs. optimal statin prescriptions for our patient cohort. Finally, we will discuss how multi-objective optimization can be used to augment the decision-making process of prescribing physicians in a way that leverages the benefits of big data and machine learning while also supporting patient-centered care.

5.1. Comparison between standard-practice (human) and machine-learning predictions

As discussed in the Introduction, the standard-practice approach to minimize SAS takes a few important patient variables into account when prescribing statin treatment. Alternatively, here we propose using machine learning that uses 70 clinically available individual patient variables as predictors (supplement 1) in a machine-learning algorithm to support prescribing care providers.

Table 3 compares predictive performances between standard-practice and machine-learning approaches. In the standard-practice group, we chose a few important individual patient characteristics, including age, comorbidity (both Elixhauser [43,44] and Charlson Comorbidity [45,46] Index), and diabetes with other chronic complications to consider as representative examples. In the machine-learning group, we examined a generalized linear model (logistic regression; 27 variables selected by forward feature selection are also summarized in supplement 1) and an artificial neural network model that used all 70 individual patient characteristics (with regularization) as examples.

Table 3.

Comparison of many different performance measures to predict SAS between a few and multiple patient data elements.

Model Name		ROCAUC	PRAUC	Accuracy	Recall	Precision
Standard practice uses a few important patient data elements	Age	59.29	20.10	56.00	57.54	18.37
	ElixScore	56.89	19.50	68.75	37.7	20.15
	CharlScore	58.63	19.68	70.11	37.85	21.24
	Diabetes with other chronic complications	51.56	15.56	82.82	7.34	23.63
Machine learning uses multiple patient data elements	GLM-Multiple	66.94	26.89	64.72	59.66	23.09
	Artificial Neural Networks-Multiple	65.04	63.86	61.38	55.57	62.88

In Table 3, we use ROC-AUC, PR-AUC, Accuracy, Recall, and Precision to show positive and overall predictions comprehensively. When selecting an appropriate model, there are three criteria to consider: (1) good in positive prediction since we aim to predict SAS and discontinuation more accurately, (2) good prediction in general or overall, and (3) stable and good prediction in general. For example, ROC-AUC is an excellent measure to understand the overall prediction. On the other hand, PR-AUC is a good measure for understanding positive prediction, essential to accurately predicting SAS. These two types of AUCs are produced across all thresholds to better evaluate overall model performance. Instead, Accuracy, Recall, and Precision are subject to the selected threshold (we use the default threshold of 0.5). Values of these predictive measures can change when the threshold is adjusted. On the other hand, they give us a general idea about the performance in the default threshold. In general, machine-learning methods (GLM and artificial neural network) that use multiple data elements have higher ROC-AUC and PR-AUC. Furthermore, the artificial neural network method has more stable and higher predicted performances among all measures. Based on the above criteria and considerations, the artificial neural network that uses multiple variables is a good choice of model.

5.2. Comparison between original and optimal prescriptions

We next compared SAS and statin discontinuation risks between original statin prescriptions for our patient cohort (extracted from the OLDW) and optimal statin prescriptions. We examined multi-objective and single-objective (SAS or discontinuation) optimizations within the optimal prescriptions to produce PSTP. Specifically, in Table 4, we apply the same prediction model to each column. We then optimize the treatment plan using the probability predicted by the artificial neural network. In the ‘original prescription,’ a patient takes drugs based on the treatment plan initially presented in the data. On the other hand, the same patient takes treatment plans based on Eqs. (1) and (2) for single and multi-objective optimization for the other three. In other words, in this simulation setting, the same patient takes four treatment plans simultaneously. Then we repeat this process for every patient and finally have four averaged group results for SAS and discontinuation. The main reason to do this is to get rid of heterogeneous bias typically presented in clinical trials to understand associated risks due to the difference in treatment plans.

Table 4.

SAS and Statin Discontinuation Risk for Original Prescription and Optimal Prescription (Produced Using Either Multi- and Single-objective Optimizations).

	Original prescription	Optimal prescription
		Multi-objective optimization	SAS single optimization	Discontinuation single optimization
SAS risk	14.82%	14.31%	10.87%	14.92%
Discontinuation risk	60.95%	55.27%	62.75%	55.11%

Compared with the original prescription, PSTP produced by SAS single optimization shows significantly lower SAS risk, and so does Discontinuation single optimization. (Table 4; highlighted in grey). In addition, PSTP produced by multi-objective optimization shows significantly lower (P < 0.00001) risk in both SAS and discontinuation when compared with the original prescription.

5.3. A proof-of-concept personalized statin treatment planning decision plot to support proactive decision-making

We next present a proof-of-concept plot to illustrate personalized proactive decision-making for statin treatment planning. Briefly, the 1-year risks of both SAS and discontinuation of all available treatment plans (Table 1) for a simulated individual were estimated using the artificial neural network prediction model, then plotted (Fig. 3). Treatment plans with lower values for both risks are located closest to the lower-left corner of the plot. In other words, this figure can help a care provider quickly identify the treatment plan with minimal risks. The prediction model determines the locations of each dot, and then the multi-objective optimization determines the PSTP (green dot). Because of inter-personal variation, each individual can have a different plot to support proactive decision-making.

Fig. 3.

Example of a personalized statin treatment planning decision plot.

X and Y-axis represent 1-year SAS and Discontinuation Risks of each treatment plan for a simulated patient. The first two letters of a treatment plan represent a statin agent and is followed by dosage. For example, SI_40 (red dot; Simvastatin 40 mg) represents the original treatment plan for the simulated patient. Among all possible treatment plans, the PSTP (identified by using Eq. (2)) for this particular patient was AT_80 (green dot; Atorvastatin 80 mg) with SAS and discontinuation risks of about 10% and 55%. This PSTP is the optimal balance between SAS and discontinuation.

5.4. Personalized patient-centered statin treatment planning decision plot to support proactive decision-making

As described earlier, we can add an individual’s preference as a further constraint in the optimization approach. In other words, we can both optimize the statin prescription and support personalized, patient-centered decision-making. We present here an example of such a decision plot for a simulated patient. In this example, because of concern regarding the potential for increased adverse effects, we wanted to avoid high-dose statin therapy for this individual. Therefore, we included this concern as a constraint in our prediction model to optimize the statin prescription in a patient-centered manner (Fig. 4). In this example, FL_80 is the preferential PSTP because this statin treatment plan is moderate intensity and the maximal balance of low SAS (~7.5%) and discontinuation (~60.5%) risks.

In this case, the constraint is avoidance of high-intensity statin (Eq. (3)), so the system only generates recommendations for low- to moderate-intensity statin. SI_40 (red dot; Simvastatin 40 mg) represents the original treatment plan for this simulated patient, and FL_80 (green dot; Fluvastatin 80 mg) is the patient-centered PSTP.

6. Discussion

In this report, we demonstrated the feasibility of a proposed machine-learning model to produce a proactive PSTP strategy that aims to prevent and minimize SAS and statin discontinuation, thereby improving the usage of statin therapy. Briefly, the model predicts a treatment plan (PSTP) with the lowest risks of both SAS and statin discontinuation, based on an individual patient’s characteristics, prior to the initiation of statin therapy. In contrast, traditional reactive strategies, such as reducing dosage or changing medication, are used to manage SAS after they manifest in current clinical settings. This reactive approach can involve trial and error in changing treatment plans, leading many patients to discontinue statin blood cholesterol-lowering treatment. Indeed, a large Internet survey [5] found that SAS was the primary reason patients discontinued statin therapy.

The concept of a proactive strategy for statin prescribing has been widely accepted by prescribing care providers, who usually take a few important personal variables into account when prescribing statins. This report’s contribution is to propose a way to scale the use of personal variables in statin prescribing by well-utilizing machine learning and big data. Although considering multiple variables (>70) can reduce more risks than the few important variables typically considered by prescribers, taking more than 70 variables into account poses too great of a burden to humans. Therefore, a need exists to use a machine-learning tool to empower prescribing care providers, and our ultimate goal is to improve this tool to support proactive decision-making continuously. Another important reason to use the machine-learning approach is that it may be easier to adapt to different patient subgroups. Although we demonstrate proof-of-concept methods for the ASCVD subgroup, these methods are not ASCVD specific. Because of the nature of machine learning, one can quickly produce and adapt models for different subgroups, such as applying the proposed method to other data cohorts and tuning parameters.

We propose using a “personalized decision plot” (as presented in Fig. 3 or 4) to empower prescribing care providers. This “personalized” plot immediately distinguishes the best treatment plan for an individual. Such visualization of the optimal versus the standard “best practice” is directly actionable by the treating care provider. Consequently, the approach we propose will create a new means of optimal decision support. Further, we can include the individual’s preference as a constraint to the personalized decision plot (such as presented in Fig. 4) to explore and demonstrate the use of this plot to support patient-centered statin treatment planning decision-making. Therefore, we argue that this kind of model, if trusted by care providers and patients, could provide significant value to support precision-health and patient-centered decision-making and, subsequently, improve long-term statin adherence.

A critical task in this proof-of-concept study is to produce a platform/process/pipeline that can flexibly accommodate different needs and use other models, e.g., in our case, we use either GLM or artificial neural networks model. In addition to these two prediction models, the pipeline also allows us to use many other prediction models, such as SVM or tree-based methods. In addition to more stable performances (i.e., Table 3), there are two other important reasons to include the use of artificial neural networks model in this pipeline: (1) artificial neural networks can learn complex relationships among variables, and (2) the model is a scalable solution, which means performance may improve when more data become available. These properties perfectly fit the progress of our development of improved data extraction and software pipeline development methods. This proof-of-concept study aims to demonstrate the feasible use of integrating prediction, optimization, and an individual’s characteristics to generate a decision plot to assist proactive prescribing decisions. For example, in Fig. 4, the decision plot is generated based on the individual’s characteristics, prediction [Artificial Neural Network (ANN) prediction, Fig. 2], constraint to avoid high dosage treatment plan, and optimization (Eq. (3)). We anticipate this constraint information comes from a conversation between physician and patient. On the other hand, Fig. 3 is a decision plot without discussion between a patient and physician (or when all treatment plans can be considered). It was produced using Fig. 2 (ANN prediction) and Eq. (2). Finally, Figs. 3 and 4 indicate the risk of 1-year SAS and discontinuation of feasible treatment plans. One may argue about the weight and balance between SAS and discontinuation, the choice of prediction and optimization model (e.g., Pareto optimization), the need to include more outcomes, the need to determine different constraints or the time to minimize risks. As a result, there are many places to improve the proof-of-concept pipeline further based on the specific user needs and environment. However, they are out of the scope of this manuscript.

The artificial neural networks tend to be poorly calibrated, and, hence, the ability for result interpretation [i.e., predicted probability (calibrated) instead of predicted score] is more limited. On the other hand, we used optimization to identify the treatment plan with minimal risk (or most confident on the prediction is close to zero harm) using either a predicted score or calibrated predicted probability (good interpretation). We note that one cannot have a label of the best treatment plan or ‘the optimal’ since the data are retrospective. In this project, we trained two sets of models, one for SAS and the other one for discontinuation. Each set was applied to 27 treatment plans and, subsequently, we have two sets of 27 predictions, whose score represents the confidence of the closeness between prediction and ground truth [39,40]. Subsequently, we use optimization to find the treatment plan with the highest confidence (optimal treatment plan). Therefore, ‘the best treatment plan’ is the treatment plan with the highest confidence (lowest predicted risks of SAS and/or discontinuation). From a multi-objective optimization point of view, the ideal PSTP is the treatment plan with the lowest risks of both SAS and statin discontinuation. In some instances, the PSTP identified may not exhibit the lowest risks for both SAS and discontinuation (see, for example, Fig. 4). Both risks are very relevant but not equal [5]. From a practical standpoint, some patients with SAS continue statin therapy, and some patients without SAS discontinue their statin therapy. A trade-off decision between SAS and discontinuation is necessary when optimizing. Therefore, we identify the PSTP as the optimal balance between both risks.

The purpose of multi-objective optimization is to simultaneously optimize several important considerations, such as benefit, risk, cost, and individual preference. In this study, we only demonstrate optimization for risks (SAS and discontinuation) and the potential to include an individual’s preference. This work does not include the benefit of LDL cholesterol or ASCVD event reduction but is essential in the statin prescription decision. Compared to SAS and discontinuation, LDL cholesterol lab test values are more challenging to obtain for four reasons: (1) much more missing values for different reasons, (2) time points and periods of the value largely vary, (3) the need to compare across EHR and Claims to accurately capture available LDL test values, and (4) the need to include different scales of the LDL value in multi-objective optimization. ASCVD events were also more challenging to obtain due to the need for longer-term follow-up. The resources and time to handle all these issues to get appropriate training data were unavailable for this study. Notably, SAS, discontinuation, and individual preference were sufficient to demonstrate an effective method that can be used for optimizing multiple targets. For translation and clinical implementation purposes, in our next planned step, we will include the LDL reduction and other practical objectives, such as costs in the multi-optimization model. In addition, we will improve the approach to handle trade-off issues among multiple targets, such as weighting of multi-objectives in various scenarios and determination of hard and soft constraints in optimization

This proof-of-concept study has some limitations. First, as a claims-based study, we could not access clinical notes information where minor SAS would be documented. Therefore, we only used ICD-9-CM codes for SAS, which are, in general, more severe. As a result, our PSTP only considers preventing or lowering the risk of more severe SAS. Second, we limited the selected cohort used in this study to newly diagnosed ASCVD patients seeking statin treatment for secondary prevention of ASCVD (an important patient subgroup discussed in the ACC/AHA 2013 and 2018 guidelines on blood cholesterol treatment to reduce ASCVD risk [1,2]). Third, to quickly test our method and understand the PSTP impact on the first statin prescription, we only selected patients who did not change their statin treatment plan during the study period. Fourth, we did not use weights to control the conflict of emphasis between SAS and discontinuation risks when producing the PSTP. Instead, we presented the personalized decision plot based on equal weight to demonstrate our method. We note that one can incorporate these weights as needed, but how to decide them is out of the scope of this work. Fifth, we only used diagnosis code to determine ASCVD in this proof-of-concept study, limiting the type of SAS we can use in this study. Sixth, in this claim data study, fill data do not provide the best resolution to determine therapy discontinuation, resulting in under-utilization of the medical therapy (e.g., forgetting to take a statin). Seventh, we do not include confidence intervals in the decision plots (such as Figs. 3 and 4) because the results are individualized. Using simulation to produce these data is a potential solution, but a detailed discussion is out of the scope of this study. Lastly, we developed our model using records for patients of ages 40 years and older. In our preliminary observational study using OLDW data, we found that SAS and discontinuation vary with age. This finding is confirmed by clinical studies and guidelines (such as [29]). These collective findings indicate that we need to develop an age-specific model in our next steps.

In future work, we will investigate the drug-drug-interaction concept and develop variables to represent it in our model best. In the current model, we only demonstrate the use of the PSTP as it relates to the first statin prescription. From a practical point of view, it would be essential to take previous statin drug failures into account to optimize the next PSTP (or previously described statin re-challenge). In addition, instead of considering only the two essential risks of SAS and statin discontinuation, we will work to add the benefit of blood cholesterol level reduction to our model for producing a more comprehensive prescription decision. We would also like to explore the use of decision curve analysis to quantify net benefit and handle weighted benefits and weighted harms. In addition, a new class of non-statin drugs, PCSK9 inhibitors, has recently shown promise for lowering blood cholesterol levels. Because PCSK9 inhibitors were approved in late 2015 and are currently very expensive, only a small number of patients in our claims data have taken this new drug to date. We anticipate that its costs will drop in the next few years, and therefore more patient records with PCSK9 treatment plans will become available as the drug is adopted. When sufficient data exist, we will include PCSK9 clinical efficacy and costs in our optimization model. Another important work is to develop a shared decision-making approach to integrate into the EHR that will incorporate the perspectives of both physicians and patients. Finally, we will also explore the multiclassification model. The literature discusses SAS as an important reason for discontinuation, and, therefore, both outcomes share some predictive information. A multiclassification model may pick up and utilize that shared predictive information and further improve performances.

Supplementary Material

Appendix A. Supplementary material

Supplementary data to this article can be found online at https://doi.org/10.1016/j.jbi.2022.104029.