Decision support systems in healthcare: systematic review, meta-analysis and prediction, with example of COVID-19

1.   Introduction

Decision-Making Support Systems (DMSS), also called Decision Support Systems (DSS), are information systems designed to assist decision-makers and interactively support all phases of a human decision-making process [1]. Intelligent Decision Support Systems (i-DSS) are DSS based on Artificial Intelligence (AI) tools to increase the impact of management support [2]. An i-DSS is a sub-discipline of traditional DSS that incorporates techniques to supply natural intelligence and uses the power of modern computers to support and enhance decision-making through machine learning [2]–[5]. The i-DSS may, for example, “respond quickly and successfully to new data and information without human intervention, deal with perplexing and complex situations, learn from previous experience, apply knowledge to understand the environment, recognize the relative importance of different elements in the decision, incorporate the knowledge of domain experts, recommend action, and/or act on behalf of the human (by a predefined authorization of the decision-maker)” [1]. In the clinical domain, DSS are denominated Clinical Decision Support Systems (CDSSs). They are defined as any electronic or non-electronic system designed to aid directly the clinical decision-making, in which characteristics of individual patients are used to generate patient-specific assessments or recommendations that are then presented to clinicians for consideration [6]. CDSSs are also defined as computer programs based on evidence-based clinical guidelines with or without AI, and designed to support healthcare providers in identifying problems, resolving them, and reducing errors [7]–[10]. With the important volume of Electronic Health Records (EHR) data along with the increase in computational power, storage and memory, it represents a great opportunity to apply Machine Learning (ML) methodologies and perform a wide range of complex tasks with impressive accuracy in the healthcare context [11].

CDSSs are based on large amounts of clinical and biological data, and then use algorithms for matching patient conditions and phenotypic patterns to the most relevant recommendations, to provide the most pertinent information and alternative actions, at each step in the clinical decision-making process. It is designed to provide practitioners with optimal, individualized and real-time support, using compiled evidence so that each patient receives efficient, effective, and customized care [12]. These systems are commonly applied in healthcare processes, such as early detection of diseases and predictive medicine providing clinicians with information about individuals at risk, disease onset and how to intervene. CDSS are also used for triage, identification of changes in health symptoms, extraction of patient data from medical records, in-patient support, evaluation of treatment, and monitoring [13]. When planning to implement CDSS, researchers and developers have to focus on the five ‘rights' of CDS: the right information, the right person, the right intervention format, the right channel and the right time in workflow [14].

Given the importance of new technologies and data analysis in reducing costs and improving the quality of healthcare, along with the important role of DSS and data mining, it is proposed some AI algorithms, their advantages, challenges and applications in healthcare. Many narrative, systematic and scoping reviews have been realized in CDSS, for diseases, either to focus on the disease like covid-19 [15],[16], cancer [17],[18], sepsis [19],[20], etc. or on the ML techniques applied [21]–[23]. And generally, researchers studied one or several CDSS purposes, like monitoring, prediction and recommendation systems. However, there is a lack of scientific research with a comprehensive view of the types of CDSS algorithms and techniques in healthcare including monitoring, alert, prediction and recommendation systems.

In this paper, we have conducted a survey of recent state-of-the-art research on CDSS using a systematic review. We used PRISMA to search and select 231 research articles based on our inclusion criteria. We included research that present ML techniques and their applications in clinical systems. We used KH coder to analyze included research, then we outlined and grouped CDSS according to their types, applications and some of their purposes.

2.   Methods

2.1.   Systematic review

We focused our systematic review on decision support systems used in healthcare. A systematic review is “a review that has been prepared using a systematic approach to minimizing biases and random errors” [24]. It is different from the traditional narrative review by its greater transparency and the reproducibility of the search [25]. In order to emphasize clinical support tools, we used the PRISMA statement [26] which is an evidence-based minimum set of items (27-item check-list) for reporting in systematic reviews and meta-analyses.

2.2.   Search strategy

A systematic search of the literature was performed in Mai 2022, and updated in September 2022, using Web of Science, PubMed, Scopus, and Cochrane. The following terms and operators were used for the search: (diagnosis system OR clinical decision support system OR clinical support tool OR computer-aided diagnosis OR health monitoring) AND (life threatening disease OR health) AND ((prediction OR probabilistic OR critical OR decision OR recommendation) AND (system OR model OR approach OR algorithm))

2.3.   Study selection criteria

Titles and abstracts were screened for eligibility. Many retained records were not coincident with the request target. So, we included primary research studies that focused on diseases, caregivers, or healthcare professionals; using artificial intelligence-based tools either for diagnosis, monitoring or prediction. We excluded studies made before 2019.

2.4.   Data collection process and data items

We extracted data from the selected reports according to the eligibility criteria. Then, we studied them one by one to find the major categories of DSS in healthcare.

3.   Results

3.1.   Main findings of the review search

The database search retrieved 1605 citations (Figure 1). After title and abstract screening, 837 articles were excluded. We screened the full texts of the remaining 448 articles. After the full-text screening, 286 articles were excluded. Then we kept only studies since 2019, leaving 231 articles. The keywords which were present in most of the selected papers were “clinical decision support system” and “artificial intelligence”, which are emphasized in Figure 2.

3.2.   KH coder

KH Coder is an open source software for computer assisted qualitative data analysis, quantitative content analysis and text mining. It can be also used for computational linguistics, and for factual examination co-event system hub structure, computerized arranging guide, multidimensional scaling, and comparative calculations. Coder allows to identify themes in large unstructured datasets, and supports several kinds of searches with frequency tables indicating what kind of words appeared frequently [27]. In order to investigate included records themes, we analyze the 231 articles using the KH Coder text mining tool. Table 1 shows frequently occurring words. The most frequent word was “support”. Frequent keywords related to our research were “support”, “decision”, “clinical”, “use”, “health”, “machine”, “care”, “learning”, “base”, and “monitoring”.

Figure 3a,b show network diagrams of words and word connections. In Figure 3a, looking at the word arrangement, “study”, “health”, “decision”, and “support” were almost in the center. In Figure 3b, “support” is a keyword because it is the center of strong co-occurrences. As for “support”, the connection of decision systems was seen mainly from “clinical”, and “decision” that co-occurred with “support”. In the co-occurrence network, knowledge extraction was classified into four groups. From the set of extracted words, group 1 was interpreted as “support”, group 2 as “decision”, group 3 as “monitoring” and group 4 as “prediction”.

Hierarchical cluster analysis (Figure 4) was performed with the minimum number of occurrences limited to 9 or more. “clinical”, “decision”, and “support” were included in the same cluster, and other terms concerning decision support systems regarding “machine”, “health” and “prediction” were also convincing. In the cluster analysis, knowledge extraction was classified into five clusters. From the clustering of knowledge extraction, cluster 1 was interpreted as “support”, cluster 2 as “monitoring”, cluster 3 as “learning”, cluster 4 as “prediction”, and cluster 5 as “diagnosis”.

Figure 3 shows a diagram where similar words are clustered and classified in a hierarchical structure, in order to hierarchically capture combinations of words having similar appearance patterns from the extracted words.

Figure 4 is focusing on the word “support” in the cluster represented in the self-organizing map of Figure 5, and “decision”, “evaluation”, and “implementation” were formed in the same cluster.

3.3.   CDSS

The CDSSs are based on the five “right” concepts representing the right information, the right person, the right intervention format, the right channel and the right time in workflow [14] (Figure 6).

CDSSs intervene in health informatics processing pipeline (Figure 7) and provide several modes of decision support, including alerts, reminders, advice, critiques, and suggestions for improved care. They are intended to assist physicians and other medical professionals in making informed decisions about a patient's care, based on electronic health records, medical literature, and clinical guidelines. We classified CDSSs retrieved from our review according to the type of the system. We sort them into alert systems; monitoring systems; recommendation systems and prediction systems.

3.3.1.   Alert systems

Computerized alert systems are widely used in clinical medicine for different specialties such as oncology [28], neurology [29], endocrinology [30], acute medical care [31],[32], etc. These systems can be categorized as simple alert systems, and alert systems combined with monitoring, recommendation or early detection systems according to the physician's needs. Simple alert systems were used for various diseases like covid-19 [33], cancer [28], diabetes [30], and influenza [34]. Some of them were used in intensive care [31] and polypharmacy [35] to improve process efficiency. Other more complex alert systems, used for dual purposes, like combined with monitoring systems in polypharmacy [36], or with recommendation systems for drug prescription [37],[38] or early detection systems for brain [29] and kidney [39] injuries or seasonal allergic rhinoconjunctivitis [40] (Table 2).

3.3.2.   Monitoring systems

Medical monitoring systems are devices or software applications that are used to track and record various physiological parameters of a patient. They are involved in several specialties like cardiolody [44], geriatry [45],[46], to help clinicians to screen patients' vitals both in hospitals [47] and remote health either to improve hospital-homecare transition [48], or simply to track patients vitals and preventive healthcare [49]–[51] (Table 3).

3.3.3.   Recommendation systems

Recommendation systems are particularly used for drug prescription [60],[61], laboratory tests [62] and polypharmacy management [21],[63], besides diseases like diabetes [64], rare diseases [65], blood pressure, thyroid [66], etc. (Table 4).

3.3.4.   Prediction systems

The prediction systems are particularly used for complex diseases with high mortality rates like cancer [72]–[74], sepsis, typhoid, etc. or diseases having a lot of ambiguity in their diagnosis, that need a good expertise to be well treated. They are used for diagnosis, and particularly for early detection of diseases. Moreover, they are used to predict the outcome of some treatment and mortality rate given some conditions [75]. During the covid-19 epidemic, these systems were used for early detection [76]–[78], vaccination [79],[80] and mortality prediction [75] (Table 5).

3.4.   CDSS algorithms

We can define two categories of CDSS, Knowledge-Based and Non-Knowledge-Based. The Knowledge-Based category uses knowledge-bases to infer the results. A knowledge base represents rules (“if-then”) and queries inside an inference engine. While Non-Knowledge-Based CDSSs depend on Machine Learning (ML) to analyze patient data. ML uses a wide number of techniques from simple data analysis and pattern recognition to fuzzy logic and inference. In this section, we are focusing on the last category and its algorithms. The most popular classification algorithms are logistic regression (LR), support vector machines (SVM), naive Bayes (NB), k-nearest neighbors (kNN), linear discriminant analysis (LDA), artificial neural networks (ANN) with deep learning techniques, decision trees (DT), ensembles like random forests and gradient boosting [94]. Besides these techniques, there are several other algorithms used in CDSS, that will not be detailed in this paper like Bayesian networks [95],[96] used in various diseases studies like covid-19 [97] and sepsis diagnosis [98].

3.4.1.   Logistic regression

Logistic regression (LR) is a statistical method for binary classification that can be generalized to multiclass classification. It has a very good performance and gives linearly separable classes [99]. LR is commonly used in healthcare due to its easy interpretability, having been used for several diseases like diabetes [64], covid-19 mortality prediction [75], septic shock early detection [86] and admission management at emergency department [100]. The probability p of an event to occur given the value x of an observable variable is calculated from the logistic function:

The quantity log[p/(1−p)] is the logarithm of the odd ratio, also known as log-odd or logit. The odd ratio of an event is the likelihood that this event will occur. It can be seen as the ratio of “successes” to “non-successes”. Technically, odd ratio represents the probability of an event divided by the probability that this event will not occur [101]. This probability can be calculated from the odd ratio odd as p=Odd1+Odd.

3.4.2.   Support vector machines (SVM)

SVMs are considered as accurate and highly robust. The objective of SVMs is to find the most suitable function of classification to separate the classes in the training data when undertaking the two-class learning task. Namely, finding the most extreme margin of hyperplanes with the best use of generalization is the motivation that drives investigating SVMs [99]. SVM algorithm is widely used in CDSS, such as for covid-19 [33] and ECG arrhythmia detection [102], and remote health [45],[49], since it permits the most appropriate classification performance on training data, perfectly classifies future data, and thereby offers the best generalization ability [103]. The function to maximize, in order to actually find the maximum margin is:

where t is the number of training examples, and αi,i=1,...,t, are non-negative numbers such that the derivatives of L_p with respect to α_i are zero. α_i are the Lagrange multipliers and L_p is called the Lagrangian operator. In this equation, the vectors w→ and constant b define the hyperplane.

3.4.3.   Naive Bayes

Naive Bayes (NB) classification method is based on simple probability using independent class assumption. The advantage of such approach is that error value would be minimum with high speed of process if number of data is big. The probability P(x) of data with unknown class is:

where c is a specific class hypothesis data; P(c|x) represents the probability hypothesis C based on x condition (posteriori probability); P(c) is the probability hypothesis c (prior probability) and P(x|c) represents the probability x based on hypothesis c condition [93].

3.4.4.   K-Nearest neighbors (kNN)

KNN is a classification algorithm based on proximity data. It uses a memory-based technique and samples of trial have to be kept during run time in this memory [93]. Considering the first sample and the second sample, then distance between the data would be calculated based on their proximity. To do so, Euclidean Distance formula is used:

where x represents testing data and y training data. dist (x,y) is the distance between data training and testing and n is the dimension of the free variable data. This classification algorithm is also used in several CDSS particularly for heart disease detection [102],[104],[105].

3.4.5.   Linear discriminant analysis (LDA)

LDA is a generalization of Fisher's linear discriminant analysis, generally used to find a linear combination of features, which separates two or more classes of data. Linear classifiers are supervised learning models used to make prediction decisions for a given observation based on a linear predictor function combining a set of weights with the components of the observation feature vector [106]. LDA is widely used for early detection of diseases using EHR data [106]–[108].

3.4.6.   Decision trees (DTs)

DTs are used in several CDSS research like diabetes [64] and covid-19 detection [75], urinary infections studies [69], and septic shock early detection [86]. They have a structured hierarchy like a tree, with internal nodes having exactly two outgoing edges. The splits are selected using towing criteria and the tree obtained is pruned by cost-complexity pruning [109]. Classification trees have target variables taking a discrete set of values while regression trees target variables use continuous values. DTs are inefficient if they learn from incomplete data, however contrarily to ANN, they are not black-box models and provide an easy interpretation of the classification system [110].

3.4.7.   Random Forest (RF)

RF used also in various disease studies like diabetes [64] and covid-19 detection [75], septic shock early detection [86] and heart disease detection [102], is based on DT as the main classifier, and on the use of ensemble learning technique to characterize information. The ensemble technique combines indicators from multiple trained classifiers to classify new instances. A random forest is a type of classifier that consists of tree-organized classifiers. Each tree makes a unit vote for the most well-known class. A random vector is independent of the former random vectors with the same distribution, and the training test is employed to create a tree [99].

3.4.8.   Fuzzy logic classifier

Fuzzy systems are used to describe uncertain phenomena, since real world is a complex system with a lot of interacting elements at different levels. The use of fuzzy expert systems based on fuzzy logic to handle uncertainties generated by imprecise, incomplete and/or vague information, can transform information gathered by experts into knowledge, which can be used later for early diagnosis of diseases or for developing treatment plans for elders [111] and patients suffering from diabetes [112] and chronic kidney disease [113]. The fuzzy classifier has the advantage of interpretability compared with other techniques like ANN.

3.4.9.   Artificial neural networks (ANNs)

ANNs are used in CDSS for prediction, diagnosis, classification and regression problems. They are based on a nonlinear approach to solve a given problem where the relationships between inputs and outputs are known. ANNs are inspired by the human brain. They present a set of processing units that work in the same way as the human brain performs its functions [114]. Data are divided into a training dataset and testing dataset, and ANN acquires knowledge through the training process and learns specific patterns from the inputs provided in the training dataset. When the network is trained, we test the obtained model on the testing dataset and evaluate its performance. Since nowadays, the processing speeds of computer systems have increased immensely, ANNs have been widely used in CDSS in many clinical studies [114] related to covid-19 detection [76], diabetes glycemic control [115], heart diseases [116],[117], liver disorder [118] and suicide attempt prediction [119].

4.   Discussion

AI is currently widely used in CDSS to improve the diagnosis, prognosis and treatment of a particular pathology. It predicts the risk for a certain disease or the probability of a medical outcome, based on various data types, from different data sources like HER, research papers, etc. Our study showed that according to medical professionals needs, several CDSS algorithms can be used for the same purpose. To predict if a patient has some disease, different classification tools can be applied, such as SVM, decision tree, random forest, etc, and data scientists have to choose the right tool with the highest accuracy and the lowest error rate. Therefore, several studies in our review were based on different algorithms for covid-19 early detection. [76] used deep learning method to analyze X-ray images and detect covid-19, while [77] proposed an ANN-based CDSS based on EHR of previous patients to help doctors in their diagnosis and [78] used linear and logistic regression to develop a covid risk calculator. All these proposed techniques aim to enhance diagnostic accuracy and assist physicians to make better-informed decisions and thereby improving patients' outcomes and increasing their satisfaction. Moreover, we found that some AI techniques like SVM, can be involved in different CDSS types. It has been used in the alert systems for covid-19 detection [33], in prediction system to help obstetricians to make more informed decisions about delivery mode [88], as well as in remote health monitoring in clinical trials [49].

5.   Conclusion

In the medical domain, in most countries, particularly in developing ones, diagnosis and treatment are based on the experience of medical staff, particularly of physicians, rather than on exploiting all the knowledge acquired in EHR data. However, recently, AI has been introduced in biomedical applications such as clinical decision support systems, diagnosis, prediction and monitoring systems. These systems have been applied in several specialties like infectious diseases, chronic diseases, and cardiovascular diseases, as well as in oncology and psychiatry. This paper presents through a systematic review using the PRISMA statement, the state of the art of CDSS and their various applications. The goal of this paper is to provide an overview of how the DSS can benefit healthcare and biomedical applications, what challenges have been faced, what methods have been proposed on the algorithm design level, and what future trends are. As a result of our study, we got a systematic review with a large sample of studies that are high quality. However, PRISMA is primarily focused on the reporting of systematic reviews and meta-analyses. So, it cannot deal with missing data, handle studies with a high risk of bias, or assess the overall quality of the evidence.

And even if CDSSs have generally a lot of benefits, risks also exist. Therefore, there are many challenges related to data safety and availability, requiring ethical clearance and data access approval, particularly with the reuse of clinical information within the system [120]. There are also challenges relative to the five ‘rights', since the CDSS must be value-added for the right target, user-friendly and useful, according to individual needs, purposes and culture. Thereby, in order to develop a good CDSS, practicing clinicians should be involved in all the CDSS phases. And further research needs to be done to improve AI explainability since medical professionals can become skeptical if they do not understand the data processing, algorithms and their outcomes [121].