Methodological issues of the electronic health records’ use in the context of epidemiological investigations, in light of missing data: a review of the recent literature

Relevant studies, without any chronological and country restriction, were identified by searching in Medline (via PubMed), Scopus, and Google scholar databases by using the search strategies presented in Table 1. After removing the duplicate studies found among the different databases, articles were manually and independently screened by both authors (TT, DP), based on their Title and Abstract and then full text reading was conducted for the final selection decision. In the case of disagreement, another scientist was asked to comment on the eligibility of the reviewed study.

Of the 1972 references initially identified from the electronic and manual search studies (PubMed: 313; Scopus: 519; Google scholar: 1140), a total of 17 studies were included in the present narrative review, which were divided in two categories:

At first, 20 duplicate records were removed, and then the remaining 1,952 records were screened based on their title and abstract. From those, 1,897 records were removed due to irrelevance to the aim of the present review. Finally, 38 records were also removed as we were not able to retrieve them from the authors after contacting them (i.e., not available in full- text). Thus, in category 1, 8 studies were reviewed, and in category 2, 9 studies were reviewed. In Table 2 the selection process of the studies is described.

In a case study published by Vuppalapati et al., [22] it was shown that selfies constitute important outpatient healthcare data which could improve the diagnosis of diseases, as well as the decision-making process. More specifically, it was reported that selfies taken for medical image purposes constitute valuable outpatient healthcare data providing new clinical insights, while they could also be used as diagnostics markers for the provision of prognosis of potential masked diseases. In addition, according to Bar-Dayan et al., [23], whose main aim was to assess the effectiveness of using the EHRs in terms of cost-savings, EHRs were shown to yield significant improvements, both to physicians, as well as to clinic practices and healthcare organizations, as they were shown to provide substantial cost- savings.

Electronic health records can assist in both the prevention, as well as the treatment of a disease. Lardon et al., [24] based on EHR data, developed rules to support diagnosis coding of chronic kidney disease (CKD) in the hospital of Saint Etienne. In another study of of Garnica et al., [25] electronic health records were shown to help in the prognosis of bacteremia, involving early diagnosis for the provision of treatments to avoid complications and death. Machine Learning (ML) techniques were applied to predict the result of blood culture for the timely administration of the correct treatment thus reducing medical costs. Furthermore, Zaballa et al., [26] presented a general framework to identify and discover the most common treatment pathways which are being exploited to treat diseases. Besides, King et al., [27] confirmed the clinical benefits of EHRs through cross-sectional data examination. EHR adopters reported benefits of EHR use in terms of clinical quality, patient safety, and efficiency, while the use of an EHR meeting Meaningful Use criteria was found to be significantly associated with reporting clinical benefits enabled by these functionalities. Except for that, as claimed by Huang et al., [28] EHRs constitute valuable tools which can help in the prediction of multi-type major adverse cardiovascular events. According to Linder et al., [29] it was also shown that EHR–based interventions can improve the smoking status documentation and increase the counseling assistance to smokers. In Table 3 the main findings regarding the contribution of the EHRs on medical quality and the health system, are presented.

In the context of EHRs, lack of documentation is mainly observed in cases when the patients do not have a symptom or comorbidity. In these cases, instead of recording a negative value for each potential symptom/comorbidity, all data fields are left missing and only the positive values are recorded. Therefore, lack of a symptom/comorbidity, lack of documentation of a symptom/comorbidity and lack of data collection regarding the symptom/comorbidity cannot be differentiated.

According to the reviewed literature, there is a variety of approaches toward managing missing EHR data; Goldstein et al., [30], who conducted a systematic review regarding the challenges faced during the development of risk prediction models based on EHRs, found that only 58 of the 90 studies (64%) evaluated addressed missing data prior to analysis. Some of the simplest methodological approaches being used, involve the selection of sub-datasets that contain complete information [31, 32], as well as the stratified mean imputation [33], while others have advanced statistical methodologies which are applicable only to continuous measures and interpolate longitudinal variables with limited individual-level variability that are typically not dependent on other covariates [34]. Despite these approaches, few studies utilized “informative observations” where the presence of a variable is meaningful for the possibly missing values [30]. Xu et al., [35] developed a deep learning unsupervised method to impute missing values in patient records and by comparing it with four other imputation techniques, they showed that the specific methodology could significantly reduce the imputation biases under various scenarios, and as a result it could empower physicians and researchers to better utilize the EHRs aiming at improved patient management.

In addition, Hwang, et al. [36] proposed a two-stage framework leading to more robust results for disease prediction based on EHRs with missing data. Two different imputation methods were implemented, the first of which replaced the missing values with the mean values of the attributes, while the second one used an autoencoder, which is an unsupervised ML algorithm. Furthermore, Wang et al. [37], based on the idea that among heterogeneous patient populations there exist homogeneous groups of patients, proposed a data driven approach for imputing the sparse patient EHRs by transferring relevant knowledge from patients with denser EHRs to their patients with sparse EHRs. In Fig. 1 an overview of the methodologies used for imputing missing data in the context of the EHRs, based on the research works included in the present review, is illustrated.

As already pointed out, some of the most important benefits related to EHRs include the easy access to computerized records, as well as the elimination of poor penmanship, which constitutes a widespread and significant obstacle in the medical world [38, 39]. Besides, EHRs provide significant cost savings, as based on the studies of Shu et al. [40] and Bar- Dayan et al. [23], it was shown that the release of EHR data to patients via smart apps can save both the hospital, as well as the patients, approximately 2 million and 1 million euros, respectively, on an annual basis. This could be attributed to the fact that, the EHR’s use can substantially reduce the redundant implementation of medical tests or the need to mail hard copies of test results to different providers [41, 42]. Additionally, several studies have also shown that EHRs, compared to hard- copies, result in reduced transcription costs through point-of-care documentation and other structured documentation procedures [43]. Furthermore, the access to electronically stored data increases the availability of data, which leads to the improvement of the ability to conduct research, as well as to the facilitation of the identification of evidence- based best health practices [44], while at the same time public health researchers by using EHRs tend to produce more beneficial for the society research outcomes. Even more, according to several studies, despite the fact that EHRs have known drawbacks when they are used solely as data sources for studies informing public health decisions [45], they contain several crucial data elements which help with a pandemic response [46, 47].

As far as the missing data handling techniques is concerned, several investigators have already tried to propose the best possible methodology, yet there is no wide consensus and acceptance in the scientific community, while there are also crucial gaps which should be addressed. As pointed out, missing information constitutes a widely spread phenomenon in routinely collected health data and often missingness is very informative and should be incorporated into the development process of prediction and epidemiological models [48, 49], as the absence of data in EHR records can substantially decrease our ability to create accurate predictions [49]. Besides, the majority of the hitherto developed prediction models are not able to provide a risk estimate when missing information exist in predictor variables, which delays their implementation and may ultimately limit guideline adherence [50]. However, the correct way of handling missing values particularly in the phase of prediction model development and in the validation dataset, solely depends on the intended use of the prediction model, and more specifically, on whether the investigator intends to allow for missing data during model application in practice [51]. So far, in clinical practise and in a real clinical setting, when applying already developed prediction models in new patients arising in the medical office to predict their risk of disease onset or disease recurrence, accounting for missing values in some of their demographic or clinical characteristics is not straightforward. Ideally, when developing a prediction model the methodology regarding the handling of missing data should be integrated, however this is not a usual case in practise, as most of the developed models do not allow for missing data [51‐63].