Integrating Clinical Knowledge and Real-World Evidence for Type 2 Diabetes Treatment

Abstract

Clinical decision support system (CDSS) plays a significant role nowadays and it assists physicians in making decisions for treatment. Generally based on clinical guideline, the principles of the recommendation are provided and may suggest several candidate medications for similar patient group with certain clinical conditions. However, it is challenging to prioritize these candidates and even refine the guideline to a finer level for patient-specific recommendation. Here we propose a method and system to integrate the clinical knowledge and real-world evidence (RWE) for type 2 diabetes treatment, to enable both standardized and personalized medication recommendation. The RWE is generated by medication effectiveness analysis and subgroup analysis. The knowledge model has been verified by clinical experts from the advanced hospitals. The data verification results show that the medications that are consistent with the method recommendation can lead to better clinical outcome in terms of glycemic control, compared to those inconsistent.

Introduction

Type 2 diabetes (T2DM) is a chronical disease that often leads to different types of complications and causes serious health problems. According to the eighth edition of International Diabetes Federation (IDF) diabetes atlas¹, in 2017, China has more than 114 million diabetes patients, accounting for 27% of the whole diabetes population world-wide. In addition to this challenging situation, China faces severe shortage of experienced physicians, especially in the large rural area. According to China Statistical Yearbook (2018)², 41% of Chinese population lives in the rural area, but only consumes 21% of national medical resources. As stated in China Health and Family Planning Statistical Yearbook (2017)³, comparisons of health workers were observed between rural and urban areas for health professionals (4.0 vs. 10.8), doctors (1.6 vs. 4.0), Registered Nurses (RNs) (1.5 vs. 5.0) per 1,000 population. In this background, a robust clinical decision support system (CDSS) for type 2 diabetes treatment is of high importance in China’s market. Given the T2DM patient’s information, the CDSS can recommend the medications and provide the evidence of why the recommendation is made. It not only assists inexperienced physicians in prescribing the appropriate medications for T2DM patients, but serves as a tool for physician education.

CDSSs have been widely applied for reducing medical errors and increasing health care quality and efficiency^4,5. There are two types of approach for CDSS: knowledge-driven and data-driven. In the context of medication recommendation, the knowledge-driven approach is to develop an expert system based on clinical knowledge (such as clinical guidelines and consensus)^6,7,8, while the data-driven one is to apply the data mining techniques on electronic health records (EHRs) to build the mapping between patient’s information and medications^9,10,11.

The advantage of knowledge-driven approach is that the recommendations are always consistent with the guidelines, which is critical to a CDSS. However, the limitation is that the knowledge provided in clinical guideline could be too general to provide fine-granular personalized recommendation¹². For example, an if-then rule for knowledge-driven CDSS could be: if one has been taking biguanides (MET) for about 3 months and his HbA1c is between 7% to 9%, then he needs to switch to dual therapy by adding another type of oral antidiabetic drug (OAD). However, the OAD includes a number of drug classes such as sulfonylureas (SU), glinides (GLN), and alpha-glucosidase inhibitors (AGIs) etc. The physician may still have the problem to decide which drug, often a product name in a drug class hierarchy, to be prescribed.

With the large amounts of EHRs, data-driven CDSS can be developed to provide personalized medication recommendation. Recently, Liu et al. proposed an algorithm to group the patients based on the similarity metric learnt from the real-world clinical data¹³. Likewise, Chen and Altman reported a Bayesian conditional probability model for recommendation of clinical orders through the data mining of EHRs^10,11. Although data-driven approach provides personalized result, the recommended results could violate the guideline, which degrades the physicians’ trust to use the system.

In this study, we propose a method to integrate the clinical knowledge and real-world evidence (RWE), for enabling the standardized (i.e., consistent with knowledge) and personalized (i.e., referring to the RWE of similar patients) medication recommendation for T2DM patients. In our approach, the clinical guideline provides the principles of the recommendation¹⁴ and may suggest several candidate drugs/drug combinations for the patients with certain clinical conditions. The RWE is generated for prioritizing these candidates and even refining the guideline to a finer level. Specifically, we represent the clinical guideline as a decision tree, in which each leaf node is a decision point with recommended modifications and the path from the root to the leaf node defines the clinical conditions of the corresponding subgroup of T2DM patients. Then the medical data are fed via the root of tree so that each leaf node is associated with a set of data samples. For each leaf node with multiple medication options, we perform the drug effectiveness analysis¹⁵ and subgroup analysis¹⁶ on the data samples, and associate the results as the evidence. These results of the leaf node can be applied to prioritize the corresponding candidate medications, or once confirmed by domain experts, to refine the clinical guideline, i.e., to further split the leaf node into a sub-tree. We have realized such an approach and developed a CDSS for T2DM treatment. We build our knowledge model based on the national clinical guideline for T2DM 2017¹⁷ and integrate it with RWE extracted from 119236 records of T2DM patients in a Chinese city. The knowledge model has been verified by clinical experts from the advanced hospitals.

Methods

Knowledge model for standardized medication

The knowledge model was built based on Chinese guideline for prevention and treatment of type 2 diabetes (2017), and had been verified by clinical experts from the advanced hospitals. An extracted rule contains information of related feature conditions and corresponding medication recommendation. According to the guideline, the glycemic control target could be set as the glycosylated hemoglobin (HbA1c) goal of less than 7.0% or fasting plasma glucose (FPG) goal of less than 7.0 mmol/L. If the target is not achieved without any therapy or after about 3 months of therapy, it should proceed to clinical treatment or adjustment of original treatment. So the treatment could be suggested by current blood glucose level, previous treatment and its duration.

Drugs commonly used for type 2 diabetes treatment include seven categories of OAD: biguanides (MET), sulfonylureas (SU), glinides (GLN), alpha-glucosidase inhibitors (AGIs), thiazolidinediones (TDZs), glucagon-like peptide-1 (GLP-1) and dipeptidyl peptidaseIV(DPP-4). Since SU and GLN have same cellular mechanism called insulin secretagogues, they can be regarded as one category of “SU/GLN”.

For each extracted rule for treatment, we set a group ID according to its corresponding therapy based on knowledge model, including Lifestyle Management, Monotherapy, Short-term Intensive Insulin Therapy, Continued Therapy, Dual Therapy, Triple Therapy, Combination Injectable Therapy. Each group contains several treatment options, as shown in Table 1.

Table 1.

Extracted rules from guideline for type 2 diabetes

Status	Blood glucose level	Previous treatment	Medication recommendation	GroupID
Untreated	HbA1c < 7.0% and FPG<7.0mmol/L	/	Lifestyle management	Lifestyle Management
	(HbA1c >= 7.0% and HbA1c < 9.0%) or (FPG<=7.0mmol/L and FPG<11.1mmol/L)	/	MET]AGIs]SU/GLN (metformin as first-line therapy)	Monotherapy
	HbA1c>=9.0% or FPG>=11.1mmol/L	/	basal_insulin+prandial_insulin]premixed_insulin] basal_insulin+any OAD	Short-term Intensive insulin Therapy
Treated	HbA1c <7.0%and FPG <7.0mmol/L	/	Continued access to this therapy	Continued Therapy
	HbA1c>=7.0% or FPG<=7.0mmol/L	Single drug, no insulin	Two drug categories with different mechanisms (with or without basal insulin)	Dual Therapy
	HbA1c>=7.0% or FPG<=7.0mmol/L	Double drug, no insulin	Three drug categories with different mechanisms (with or without basal insulin)	Triple Therapy
	HbA1c<=7.0% or FPG<=7.0mmol/L or use insulin	/	basal_insulin+prandial_insulin]premixed_insulin] basal_insulin+prandial_insulin+any OAD(except SU/GLN)	Combination Injectable Therapy

For example, information like “if lifestyle alone fails to achieve blood glucose control, monotherapy should be initiated” in guideline would be converted to a rule like: if (HbA1c >= 7.0% and HbA1c < 9.0%) or (FPG<=7.0mmol/L and FPG <11.1mmol/L), then initiate Monotherapy (MET or SU/GLN or AGIs). Metformin (MET) should be the first priority when initiating therapy.

Real-world evidence integration for personalized medication

Based on the knowledge model, the clinical guideline provides the principles of the recommendation and may suggest several candidate drugs/drug combinations for the patients with certain clinical conditions. The RWE is generated from real-world data to suggest the priority of these candidates and even refine the guideline to a finer level. Evidences mainly include feature statistics that describe a group of similar patients, usage percentage of candidate medications in the patient group and their corresponding outcome of glycemic control. When physicians face several options, RWE data (e.g. outcome or effect evaluation) become very important for identifying the optimal and personalized treatment. Figure 1 shows our methodology of integrating clinical knowledge and real-world evidence for type 2 diabetes treatment. First, we represent the knowledge model derived from clinical guideline as a decision tree, in which each leaf node is a decision point (corresponding to each GroupID) with recommended modifications and the path from the root to the leaf node defines the clinical conditions of the corresponding group of T2DM patients. In step 1 of Figure 1, the medical data are fed via the root of tree so that each leaf node (GroupID) is associated with a set of data samples. In step 2, by knowledge model, the data samples could be divided into two parts of guideline-concordant and guideline-not-concordant samples depending on its prescription, and we preform data-based evaluation of standardized medication and its relationship with effectiveness. In step 3 and 4, for each group at leaf node with multiple guideline-concordant medication options, we perform the drug effectiveness analysis¹⁵ and subgroup analysis¹⁶ on the data samples, and the results can be associated as the evidence. These evidence of the leaf node can be applied to prioritize the corresponding candidate medications, or once confirmed by domain experts, to refine the clinical guideline, i.e., to further split the leaf node into a sub-tree.

Figure 1.

Methodology of integrating clinical knowledge and RWE for T2DM treatment.

1. Applying medical data into the knowledge model and evaluating

Applying medical data into the knowledge model, we obtained seven groups. A group corresponds to a leaf node and contains data samples with similar features and different medication patterns in their prescriptions. Comparing the medication in physician’s prescription with standardized medications from knowledge model, data samples in each group can be divided into two kinds: “guideline-concordant” or “guideline-not-concordant”. Data samples were categorized to the guideline-not-concordant cohort if their prescription is not in any of standardized medications of this group. The standardization degree for medication and its relationship with effectiveness were evaluated by guideline adherence of prescriptions and the rate of achieving glycemic control goal.

To reveal the relationship between guideline-concordant medication and its effectiveness indicated by glycemic control goal achieving rate, multivariate logistic regression model was applied to adjust for confounding factors. Given that the glycemic control goal achieving rate could be affected by the baseline blood glucose level, the baseline FPG was identified as a confounding factor. Adjusted odds ratios (ORs) of achieving glycemic control goal could be obtained via multivariate logistic regression. The same analysis was applied for each group.

2. Medication effectiveness analysis and subgroup analysis

Our approach enables the guideline adherence and personalization recommendation. In each group, we first compared the effectiveness of all guideline-concordant medication options at the whole data samples of the group, where the guideline-not-concordant medications were excluded. Similarly, adjusted OR, P-value in multivariate logistic regression model were used to reveal the relationship between medication options and glycemic control goal achievement. Here, we identified confounding factors by selecting the features that were associated with the medication options, from the basic information, physical measurement and history of diseases.

If the result of the medication comparative effectiveness analysis was not statistically significant at the group level, we would apply Model-based recursive partitioning approach (MOB)¹⁶ on the leaf node to further “grow” the tree for subgroup analysis.

The basic idea of MOB is that each node is associated with a single model. The datasets are split into different subsets based on partitioning variables to make the model parameters stable and the distributions of the response values are most different. That means a precision cohort fitted well or with strong evidence may be found after further partitioning and the drug effectiveness result of the subgroup determines the final recommended prescriptions.

The based parametric model in our MOB tree is logistic regression in which medication option (pairwise or multiple options) is as independent variable and glycemic control goal is as dependent variable. For instance, binary medication options: MET or SU/GLN, glycemic control goal: FPG<7.0mmol/L.

The iterative steps of MOB method are as below:

1) fitted the logistic regression with glycemic control goal as response variable and binary medication options as independent variable on all observations. Given n observations , the model can be fitted by minimizing some objective function yielding the parameter estimate Maximum likelihood method or least squares method can obtain the solution of

$$\stackrel{⌢}{\theta }=\underset{\theta \in \Theta }{\mathrm{argmin}}\hspace{0.17em}{\sum }^n{}_{i=1}\hspace{0.17em}\Psi {\text{(y}}_{\text{i,}\hspace{0.17em}}\text{θ)},\hspace{0.17em}where\hspace{0.17em}{y}_i=\{\begin{array}{c}1,\hspace{0.17em}\text{FPG}\hspace{0.17em}\text{<}\hspace{0.17em}\text{7}.0mmol/L\\ 0,\hspace{0.17em}\text{FPG}\hspace{0.17em}\ge \hspace{0.17em}7.0\text{mmol/L}\end{array}$$

2) test the parameter instability¹⁸ on partitioning variables (age, gender, historical diseases and other demographic or medical characteristics).

3) if significant parameter instability (i.e., the smallest P values) is detected, split the sample with respect to the partitioning variable into two subsets.

4) repeat step (1) ~ (3) until there is no significant parameter instability on each node.

MOB is a recursive algorithm, since each subgroup will be further split until recursive end condition is reached.

Results

Data Set

We demonstrate the effectiveness of our method using a dataset that is a collection of type 2 diabetes patient records across multiple hospitals in a city of China with time spanning from 2015 to 2017. In total, there were 119236 records with the main diagnosis as type 2 diabetes, corresponding to 50061 unique patients. Each sample in the medical data recorded one visit of T2DM patient, which consisted of basic information of the patient, diagnosis made by the physician, blood glucose level (FPG), and the medication prescription. The distribution of age centered around 60 to 80 years old. Females accounted for 61% of the whole data. Patients with BMI more than 24 accounted for 42%. Those demographic characteristics described from data agree well with the fact that obese people are more susceptible to type 2 diabetes.

Medication Pattern Mining

Here we used association rule mining method (Aprior algorithms¹⁹) to identify the frequent sets of medications with a support threshold of 0.01%. As shown in Table 2, top 10 medication patterns identified for type 2 diabetes treatment were all consistent with the instruction from the knowledge model. Among all prescriptions, the most popular one was MET only, which accounted for about 46% of all samples. This is in consistent with the fact that MET is the first priority in T2DM therapy except for special cases. In addition, a combination of MET and SU/GLN was very common in Dual Therapy, which accounted for about 15% of all samples.

Table 2.

Description of top 10 medication patterns for T2DM

	Medication Patterns*	Med_num	Ratio
1	MET	1	46.21%
2	SU/GLN	1	16.67%
3	SU/GLN+MET	2	14.88%
4	MET+basal_insulin	2	7.82%
5	SU/GLN+basal_insulin	2	2.34%
6	AGIs	1	1.74%
7	MET+AGIs	2	1.43%
8	SU/GLN+MET+basal_insulin	3	1.15%
9	SU/GLN+MET+AGIs	3	0.99%
10	SU/GLN+AGIs	2	0.85%

^*In combination prescriptions, each drug category is connected using the “+” sign. Medication patterns showed in guideline are marked in bold.

Applying medical data into the knowledge model and evaluating result

Applying medical data into the knowledge model, we obtained seven groups. As shown in Table 3, by comparing the medication in physician’s prescription with standardized medications from knowledge model, the standardization degree for medication in each group was described by guideline-concordant number, guideline-concordant ratio and main guideline-not-concordant medication.

Table 3.

Statistical results of guideline adherence in medications for each group

	GroupID	Condition	Total	Guideline- concordant number	numberconcordant Guideline- concordant ratio	Main guideline-not- medication*
1	Lifestyle Management	Untreated, FPG<7.0mmol/L	28107	14744	52.46%	Single drug(31.83%) Double drug(14.95%)
2	Monotherapy	Untreated, 7.0mmol/L<=FPG<11.1 mmol/L	10406	3478	33.42%	No drug(44.01%) Double drug(19.48%)
3	Short-term Intensive insulin Therapy	Untreated, FPG>=11.1mmol/L	1895	271	14.3%	Single drug(38.05%) No drug(37.47%)
4	Continued Therapy	Treated, FPG<7.0 mmol/L	58195	56934	97.83%	Single drug(0.97%) Double drug(0.55%)
5	Dual Therapy	Treated, FPG>=7.0mmol/L, used single drug but no insulin	11767	191	1.62%	Single drug(97.50%) No drug(0.76%)
6	Triple Therapy	Treated, FPG>=7.0mmol/L, used double drug but no insulin	4567	16	0.35%	Single drug(7.90%) Double drug(91.48%)
7	Combination Injectable Therapy	Treated, FPG>=7.0mmol/L, used three drugs or used insulin	4299	166	3.86%	Single drug(14.31%) Double drug(62.94%)
8	All	/	119236	75800	63.57%	/

^*In guideline-not-concordant medication, only top 2 are showed. Items in bold are guideline-not-concordant ratios greater than 50%.

In summary, the majority (63.57%) of medications were concordant with the guideline. Especially in “Continued Therapy” group whose instruction is maintaining previous treatment, the guideline-concordant ratio was up to 97.83% because prescriptions always follow previous prescriptions in the data. However, in “Dual Therapy” group, the guideline-concordant ratio was as low as 1.62% since 97.50% patient visits supposed to add a new drug category according to guideline were still prescribed with previous single drug category.

There were 67493(57%) records with FPG after about 3 months’ therapy. Based on these samples we evaluated the relationship between guideline-concordant medication and glycemic control goal achieving rate, as shown in Table 4.

Table 4.

Association between guideline-concordant medication and blood glucose control for each group

GroupID	Guideline- concordant	Total number	Number of not achieving FPG goal	Rate of not achieving FPG goal	Number of achieving FPG goal	Rate of achieving FPG goal	Base FPG mean	P value*	OR*	Note*
Lifestyle Management	yes	7787	1481	19.02%	6306	80.98%	6.02	<0.001	1.20	significant concordant > not concordant
	no	4402	967	21.97%	3435	78.03%	5.95
Monotherapy	yes	1727	638	36.94%	1089	63.06%	8.28	<0.001	2.48	significant (achieving) concordant > not concordant
	no	3670	2174	59.24%	1496	40.76%	8.17
Short-term Intensive insulin Therapy	yes	148	124	83.78%	24	16.22%	14.23	0.65	0.50	no significant difference
	no	795	572	71.95%	223	28.05%	13.94
Continued Therapy	yes	36241	6629	18.29%	29612	81.71%	6.01	0.30	1.10	no significant difference
	no	733	147	20.05%	586	79.95%	6.06
Dual Therapy	yes	105	63	60.00%	42	40.00%	9.67	0.34	0.82	no significant difference
	no	7052	3557	50.44%	3495	49.56%	8.70
Triple Therapy	yes	11	5	45.45%	6	54.55%	9.15	0.60	1.38	no significant difference
	no	3055	1615	52.86%	1440	47.14%	8.84
Combination Injectable Therapy	yes	58	36	62.07%	22	37.93%	9.19	0.45	0.79	no significant difference
	no	1709	970	56.76%	739	43.24%	9.10
All	yes	46077	8976	19.48%	37101	80.52%	6.17	<0.001	3.62	significant (achieving) concordant > not concordant
	no	21416	10002	46.70%	11414	53.30%	7.94

^*Items in bold are statistically significant at P<0.05 and OR>1.

According to the evaluation result, we found that guideline-concordant treatments was associated with better clinical outcome in FPG control than guideline-not-concordant ones. The contrast of FPG-goal achieving rates show that patients with guideline-concordant prescriptions acquire higher FPG-goal achieving rate next visit time. For total samples, the FPG-goal achieving rate for guideline-concordant prescriptions was as high as 80.52%, while the FPG- goal achieving rate for guideline-not-concordant prescriptions was just 53.30%, with P value <0.001 (SS) and a rather high adjusted OR=3.62.

As mentioned before, to reveal the relationship between guideline-concordant medication and glycemic control goal achieving rate, multivariate logistic regression model was applied to adjust for confounding factors. Here base FPG was regarded as a confounding factor, which stands for the baseline level of blood glucose and has effect on FPG goal achieving rate. The mean value of base FPG was also shown in Table 4. For example, in “Dual Therapy” group, by descriptive statistic a contrary finding was that patients with guideline-concordant prescriptions acquire lower FPG- goal achieving rate (40% vs 50%). By adjusting of confounding factor, such as the base FPG (mean base FPG for guideline-concordant: 9.67mmol/L vs guideline-not-concordant: 8.70mmol/L), the adjusted P-value shows the no significant difference between guideline-concordant medication and glycemic control goal achieving rate.

For the “Triple Therapy” and “Combination Injectable Therapy” groups, the sample size is greatly unbalanced, with much fewer samples consistent with the guidelines than those inconsistent with the guidelines. The results in these groups were considered with no significant difference.

Medication effectiveness analysis and subgroup analysis result

Our approach enables the guideline adherence and personalization recommendation. First, the clinical guideline often suggests several candidate drug classes/drug-class combinations for similar patients with certain clinical conditions. Then the real world evidence would serve as the information for prioritizing multiple candidates and even refining the guideline to finer subgroups via drug effectiveness analysis. It helps physicians make informed decisions. In each group, we need to compare the effectiveness of all guideline-concordant medication options based on the data samples of the group.

Taking the group of “Monotherapy” for example, the usage number of each guideline-concordant medication and its effectiveness (FPG goal achieving rate) was shown in Table 5 with descriptive statistic. Since the number of using AGIs is quite small, we give the comparative effectiveness analysis just between MET and SU/GLN. The result shows that the adjusted P value=0.34 and OR=0.895, which means there was no significant difference in effectiveness between MET and SU/GLN. Here BMI, age and history of hypertension were identified as the confounding factors to adjust OR, because they were associated with medication options or glycemic control goal achievement significantly.

Table 5.

Descriptive statistic result of medication usage and effectiveness in “Monotherapy” group

	Total number	Number of not achieving FPG goal	Rate of not achieving FPG goal	Number of achieving FPG goal	Rate of achieving FPG goal
MET	1200	436	36.33%	764	63.67%
SU/GLN	498	191	38.35%	307	61.65%
AGIs	29	11	37.93%	18	62.07%
All	1727	638	36.94%	1089	63.06%

If the result of the comparative effectiveness analysis for medication options is not statistically significant at the group level, we apply recursive partitioning approach on the leaf node to further “grow” the tree for subgroup analysis. Here we grow a MOB tree on “Monotherapy” group to perform subgroup analysis, as shown in Figure 2. We set stopping condition: significant parameter instability P value <0.5, maximum depth of MOB tree maxDepth= 3. The partitioning variables are detected after parameter instability test, they are BMI (kg/m²) and age (year). The partitioning point for BMI is 23.34 kg/m², 58 years for age variable. As shown in Figure 2A, Monotherapy group were split into three subgroups (cluster) according to BMI and age. Three single logistic regression model fitted on each cluster, and regression coefficients and adjusted ORs of treatment options are shown in Table 6. Finally, we compared the FPG goal achieving rate using MET or SU/GLN respectively, as shown in Figure 2B. Although it is not significant due to the small number of patient samples, SU/GLN can lead to better outcome in cluster 1. In cluster 2 and 3, MET can lead to better outcome which is consistent with the guideline (Improper use of SU may cause hypoglycemia especially in elderly patients and also cause weight gain). According to the results of subgroup analysis, patients in cluster 1 (BMI<=23.34 kg/m², age<=58 years) are recommended to SU/GLN; patients in cluster 2 (BMI<=23.34 kg/m², age>58 years) are recommended to MET; for patients in cluster 3, recommendations are with no preference currently.

Figure 2.

Result of subgroup analysis with recursive partitioning approach in Monotherapy group. (A)Sub tree for subgroup splitting, (B) Glucose control rate for each medication option in all group and each subgroup.

Table 6.

Result of comparative effectiveness analysis on three subgroups (clusters) of Monotherapy group

	Adjust_OR	Coel	P value	Count	FPG goal achieving rate using MET	FPG goal achieving rate using SU/GLN	Precision recommendation
cluster_1	1.64	0.50	0.24	124	58%	71%	SU/GLN
cluster_2	0.84	-0.17	0.36	704	72%	68%	MET
cluster_3	0.96	-0.04	0.79	812	57%	56%	MET or SU/GLN
All_group	0.95	-0.06	0.53	1664	63%	62%	/

RWE-integrated CDSS

A CDSS for type 2 diabetes treatment is built with real-world evidence. After collection of patient data including basic information, lab or examine results, previous treatment, historical diseases like comorbidity or complication, the CDSS system can recommend the medications and provide the evidence of why the recommendation is made, according to the knowledge model and data-based evidence of similar patients. As shown in the top area of Figure 3, the system provided comprehensive medications for T2DM patients including blood pressure management, blood glucose management, blood lipid management and blood platelet management respectively. And each kind of medication recommendation includes class name, generic name and product name. For patients with certain conditions, there could be more than one treatment options in every management module. Physicians can switch the treatment by clicking the drop-down menu. After clicking the evidence icon next to the medication options, both the data-based evidence and knowledge-based evidence are showed at the bottom. In the bottom-right area of Figure 3, for the data based evidence in “similar patients” tab, feature statistic, using percentage and glucose control rate of medication options in similar patients are showed in the table, the pie chart and the bar chart respectively. As shown in the bottom- left area of Figure 3, the path in the decision tree is highlighted, indicating why the recommendation is made for this individual patient according to the clinical guideline. In the real application, the physicians can turn to the CDSS for 1) reasonable medication options according to our knowledge model and 2) more detailed evidence from similar patients’ data to prioritize the corresponding candidates. After choosing the medication treatment, the physician could further customize the treatment. He or she could select/unselect the drug class by the check box, and choose the drug with specific generic name and product name. Till now, we have performed the pilot testing on the CDSS system in several hospitals in China.

Figure 3.

Screen shot of the RWE-integrated CDSS

Discussion

In this study, we proposed a method and system by integrating clinical guideline and RWE for T2DM treatment. The clinical guideline provides standardized medications and RWE is generated to prioritize the corresponding candidates such as effectiveness in similar patients. For similar-patient group that is not statistically significant in effectiveness, we apply recursive partitioning approach to further “grow” the tree for subgroup analysis.

The strengths of this study include the integration of the clinical knowledge and real-world evidence, the achieving the standardized (i.e., consistent with knowledge) and personalized (i.e., referring to the RWE of similar patients) medication recommendation for T2DM patients, the medication effectiveness analysis and subgroup analysis, and finally the CDSS system realization and application.

The limitation of the current work is that evidence was based on drug categories in treatment. In fact, a physician need to decide which generic name or product name of drug can be most appropriate. Moreover, we only evaluated short term glycemic control goal achievement rather than long term outcome improvement such as the occurrence of complications. When more data are accumulated with longer time span, we can refine the recommendation and evidence to generic or product level and evaluate long term outcome improvement of the patients, so that the system will be more helpful and practical for physicians. Another limitation is that the outcome used to evaluate the glycemic control has been discretized as goal-achieving and goal-not-achieving, ordinal regression model for continuous outcome can be applied with more comprehensive evaluation.

Our approach is tested with type 2 diabetes patients’ data. The method can be generalized to other chronic diseases based on its guideline. By customization of medication pattern, features, effectiveness, and related parameters, our approach can be applied in different scenarios.

Conclusion

In this paper, we proposed a method and developed a CDSS by integrating clinical knowledge and real-world evidence for T2DM treatment. We built our knowledge model based on Chinese clinical guideline for prevention and treatment of type 2 diabetes (2017) and integrated it with RWE extracted from more than 110,000 records of T2DM patients in a Chinese city. The knowledge model has been verified by the clinical experts from the advanced hospitals. The data verification results show that the medications that are consistent with the method recommendation can lead to better clinical outcome in terms of glycemic control, compared to those are inconsistent. Furthermore, the method and system enabled the personalized medication recommendation by comparing the effectiveness of treatment options in terms of the clinical outcomes like glycemic control goal achieving rate.