The classical approach of the clinician to the management of data for clinical research is based on the premise that it is necessary to obtain quality information in order to secure reliable results that are applicable to patients. The difficulty of obtaining such information leads to attempts to optimize the process, applying prospective designs, randomization and a working hypothesis before the data are analyzed.

As an alternative to the exclusive use of data collected in an orthodox manner, Big Data Analysis (BDA) offers the novelty of detecting the underlying structure and knowledge in large bodies of information, even when the latter do not seem to be structured.

A popular definition of the concept13 is that “Big Data consists of data sets with such a large volume and such a broad structural variety that specific technology and analytical methods are needed to process them and transform them into knowledge or value”.

Specific BDA techniques have been successfully used in fields such as marketing, strategic decision making in the business world, banking, transport, logistics, insurance or the detection of fraud in electronic commerce. There are no reasons to believe that BDA cannot be applied to our setting, where we continuously make strategic or concrete decisions in given patients whose characteristics – while complex – are often repetitive.

A recent example, which we will address again further below, is the follow-up of influenza epidemics based on analysis of the searches made on the internet. In this case we use databases with many millions of registries in order to draw rapid and reliable epidemiological conclusions.14

The open code philosophy has acquired a strong presence in the world of Big Data, allowing its use without the need for major economical investments. Projects such as the Apache Hadoop15,16 include a whole range of resources in their ecosystem (HDFS, Spark, MapReduce, Impala, HBase, Hive) that allow the low-cost construction of a Big Data batch processing system (analysis of data already stored in large relational or non-relational databases) or Big Data stream processing system (analysis and processing of data as they are generated).

The concept of Machine Learning (ML) dates back to the mid-twentieth century, and was defined in an article published by Samuel in 195917 as an area of AI that uses computational algorithms and statistics to give computers the capacity to “learn”, i.e., to improve their results in a specific task after processing a sufficient volume of information and without explicit external instructions (which may be biased) from the programmer.

The scope of ML is closely related to other fields such as simulation and modeling, the optimization of systems and statistics. All of these areas make intensive use of common mathematical techniques that require specific training.

Some basic concepts need to be explained at this point. We must use a format referred to as a “data frame” to work with data in ML. A data frame is a matrix where each row corresponds to one of the patients or registries, and each column corresponds to one of the recorded variables (Fig. 2).

Figure 2.

Typical structure of a data frame, composed of rows for each of the registries, with an anonymized identifier, and columns corresponding to the values of different variables for each case. The last column shows the target variable (in this case mortality). Here we can see a section of some registries of the database used as an example in the material included in the GitHub repository.

(0.28MB).

As an example, suppose we wish to develop a model for the prediction of mortality according to the following variables: type of admission (medical, elective surgery or emergency surgery), age, sex, APACHE II score upon admission, SOFA score upon admission, and blood lactic acid levels. In this case we have as many rows as there are recorded patients, and each row has a column identifying the episode (on an anonymous basis) and a column for each of the explanatory variables that we have defined, plus another column corresponding to the target variable (in this case mortality as a dichotomic variable [YES/NO]).

For those readers who wish to reproduce the examples provided in this review, we have prepared a series of anonymized data from our own database, and have entered them in an MS Excel table in a GitHub repository (https://github.com/anunezr/revision_medicina_intensiva).

There are three main variants of machine learning:

• •

  Supervised learning. Each registry is labeled with a value of the target variable, and use is made of different techniques capable of predicting the value of that variable in a new registry. In our above example, the target variable is mortality (YES/NO). Once the system has been “trained”, it must be able to predict the value of that variable corresponding to an episode that has not been previously presented to the system. Therefore, the training data and the data of the test must be different and stored separately. This variant is used in classification, prediction and similarity detection tasks.

• •

  Unsupervised learning. In this case the aim is to detect patterns or trends in data without using a target variable. This variant is used for example to automatically classify patients into groups, and to reduce the number of variables and the complexity of the models.

• •

  Reinforcement learning. In this case the system pursues an objective or reward, and progressively learns as the environment with which it interacts is explored, and which is not known beforehand – avoiding actions with a negative reward and seeking to conduct actions with a positive reward. An example of this would be a system that learns and adjusts the empirical antibiotic regimens prescribed by the clinicians as they receive the results and characteristics of the septic patients they see.

Table 1_Sup, included in Annex A of the Supplementary Material, summarizes the different ML techniques for each of these variants. The use of such techniques requires knowledge in programming and control of the concepts used in AI, which normally are not within the reach of the clinician. Because of this, a new collaborative working approach is being consolidated, including the introduction in the Units of a new professional with basic clinical knowledge and advanced skills in statistics and the tools and methods of BDA and ML. These professionals, working jointly with the clinicians, can help us to draw value from a large amount of clinical information which right now is filed in our databases. This “man in the middle”, now already found in leading technological development centers, will soon become an important element in the teams of our Units, and will help to plan strategic decisions and optimize the software tools in order to get the most out of them.

For those readers with knowledge of R or Python and who wish to acquire basic practice in the field of ML, the GitHub repository associated to this review (https://github.com/anunezr/revision_medicina_intensiva) includes a series of scripts that use the different techniques reflected in Table 1_Sup, fundamented on the anonymized information of the database provided as an example.

Table 1 provides a “road map” for putting an ML system into practice, adapted to a concrete problem.18 Each of the considered aspects would merit a review as extensive as this article, if not more so. The aim of the table is to explain that the use of BDA and ML requires complex methodology and systematization that in turn demand specific knowledge and experience in order to obtain results. Only under these premises can we extract knowledge from the data in an efficient manner.

In order to make use of BDA and ML techniques, we must master at least some of the analytical statistical computation languages (R, Python or Java); control SQL as a database consulting tool; and know how to use the code libraries employed in this field. The Supplementary Material provides a brief description of the most widely used software.

In the face of this mix of acronyms and technicisms, intensivists tend to ask themselves whether BDA and ML techniques can really be applied in their working environment, and how to do so in an accessible way. In truth, the purpose of this review is to make the clinician aware that entering the world of AI applied to intensive care medicine is possible, but demands structural changes and investment in human resources and technology, as well as an expanded vision of the concept of clinical research.

Logically, the first requirement is a data acquisition and storage system allowing subsequent analytical processing. In Spain we are still far from achieving this, since the existing electronic clinical information systems are not universally implemented in our Units, and an inter-hospital standard is even less common. Furthermore, many of those centers that do have such applications lack efficient access to the stored data – often because the system purchasing contract did not initially contemplate this aspect, which usually comes at an added cost. These paradoxical situations, which border on ethical neglect by keeping clinical data “hostage” of the suppliers of the commercial software, must be urgently resolved if we want to work with the new BDA and ML tools for the good of our patients.

A second consideration is the need for hospitals to assign specialized staff from the Information Technology Department to the custody, processing and analysis of the clinical data of our Units, working in collaboration with the physicians and nurses in order to get the most out of the stored information.

A third consideration is the need to work with the local and regional Ethics Committees to guarantee data security and privacy, and to not obstruct opportunities for improvement of the processes and clinical care which exploitation of the stored data can afford.

Secondary analysis of electronic health records (EHRs)19 refers to processing of the clinical data of patients generated as a subproduct of their acquisition with healthcare purposes, and can be used as an aid in strategic decision making in a Unit, or on a point basis for a concrete patient.

In seeking to endorse clinical research in the intensive care setting, some institutions with very large clinical databases have placed them at the disposal of investigators throughout the world, following due anonymization to preserve data privacy. The currently most popular initiative of this kind is the Medical Information Mart in Intensive Care (MIMIC-III),20,21 the main characteristics of which are its public accessibility and the excellent quality of the filed data, comprising clinical information of all kinds, drawn from over 58,000 critical patients – both adults and newborn infants.

Another example of such databases is the eICU Collaborative Research Database,22 comprising information from a combination of many Intensive Care Units throughout the United States. The registries of the collaborative database compile data from almost 200,000 patients admitted to critical care in 2014 and 2015.

As we can imagine, although such secondary analysis may be very useful for a Department of Intensive Care Medicine or for a clinical investigator, its potential is boosted if the data from several Units can be combined to produce a large multicenter database. The effect is similar to what we see on comparing the conclusions that can be drawn from a single center trial versus a multicenter trial.

However, a number of difficulties must be overcome in order to do this.

Each Unit compiles the data using software that might not be the same as that used in the rest of the Units we seek to merge. We therefore must ensure conversion to a common data structure allowing combined or pooled analysis. This is what we referred to above as “mapping”. In this regard, ontologisms have been introduced that allow us to express and map the filed concepts in a uniform manner. Some of them, such as the Observational Medical Outcomes Partnership (OMOP)10 or Informatics for Integrating Biology and the Bedside (i2b2),23 offer tools to facilitate local database migration to a common data standard.

On only considering the studies published in recent years using the public MIMIC20 database and software libraries with latest-generation algorithms and free access, we can see that interesting results have been obtained in very diverse areas of intensive care medicine (Table 2).

Furthermore, an advantage of this strategy is that the results are fully reproducible: we can take the data set used by the authors and reproduce the statistical and ML methods they have used, even with our own information, since they are routinely available in public repositories such as GitHub.24,25 This causes the studies to be subject to additional quality control, and allows anyone to check both the methods and the results obtained.

There are conditioning elements of a legal and ethical nature that can complicate the data sharing process, and it may prove necessary for the local Ethics Committees to authorize access and check the safekeeping of data security and privacy.

The logistics of the process require coordination among the participating centers and technological and hardware support concordant with the project, involving multidisciplinary teams capable of dealing with the challenges which the different aspects of such a project require.

Once these difficulties have been overcome, the opportunities for improved management, benchmarking, and of progress in clinical research are enormous, and without doubt will compensate the efforts required to incorporate a Department of Intensive Care Medicine to projects of this kind.

Since most of the information which professionals enter in the case history is in the form of free text (also known as natural language), a possible application of AI techniques in medicine is the automated creation of structured information from free text, as well as the classification or phenotyping of patients into different groups according to the contents of the text written by the professionals. These techniques, known as natural language processing (NLP), are advancing quickly.26 There are two possible approaches:

• •

  The extraction of concepts from the text using tools such as cTAKES27 is quite well developed in the Anglo-Saxon world, though there are few practical options for texts not written in English. The research group in Data Mining of the Biomedical Technology Center of the Polytechnic University of Madrid (Spain) is working with a tool derived from the cTAKES known as TIDA,28 and in future it is possible that systems of this kind will become part of our computer-based resources.

• •

  Another alternative is the use of convolutional neural networks (CNNs),29 but this type of strategy poses an inconvenience in that the algorithms obtained are not easily interpretable by physicians – though this could become irrelevant if the model functions correctly – and the result obtained are variable and conditioned to the sought phenotypes.

Since the clinical data stored in our clinical information systems are not compiled with the same demanding criteria as in the context of clinical trials, how can the secondary analysis of patient care information allow reliable conclusions to be drawn? They key lies in the volume of the data and in the capacity of the ML tools to detect structure in the midst of caos.19 As an example, we can use follow-up of the incidence of the H1N1 influenza epidemic of 2008 based on the counting of Google searches made using the term “flu”,14 and which has continued to be validated posteriorly (Fig. 3) – exhibiting perfect correlation with the registry of cases obtained by means of more orthodox methods.

Figure 3.

One of the first examples of the application of BDA to clinical research. During the N1H1 influenza epidemic of 2008, the Google searches using the term “flu” or its symptoms proved very precise in predicting the increase in number of cases and the severity of the clinical condition. The image at top shows the number of searches, while the image at bottom shows the evolution of cases according to the Official Influenza Surveillance System in Spain (the time scales have been adjusted to ensure vertical correspondence between the two figures). This finding has subsequently been used as an epidemiological surveillance system, with great success.

(0.25MB).

Although to date it cannot be affirmed that BDA and ML have displaced the traditional clinical research methods, it is true that they allow the detection of trends or patterns in large bodies of data that may go unnoticed to the investigator, and which can guide the design of new studies adhered to the conventional methodology.30 Furthermore, they are the only alternative in those cases where it is logistically not possible to contemplate an orthodox clinical trial of sufficient statistical power to resolve a relevant question or issue.

In a recent article appearing in Intensive Care Medicine, McLennan et al.31 offered an excellent analysis of the problems which the sharing of clinical data included in databases may pose. Since the coming into effect of the European Directive on General Data Protection Regulation (GDPR)32 in May 2018, the institutions must comply with legislation referred to the handling of personal data on patients in the European Union. The equivalent of this in the United States is known as Health Insurance Portability and Accountability Act (HIPAA), of 1996.33

The key issue is data anonymization: if the information is completely anonymous, there is no legal problem in sharing it from the right to privacy and security perspective. But when can we consider the data to be completely anonymous? Only when they do not lead to unequivocal identification of the patient upon being complemented with other data (e.g., even if we do not know the name of the patient or his or her history number, if we know the date of admission and age, we may identify the patient involved). We also must avoid unequivocal identification of the health staff appearing in the case history of the patient. In all other cases the data are considered to be pseudo-anonymized, and in this situation the GDPR directive would oblige us to request patient consent to use of the information.

Database anonymization is therefore crucial, and we must standardize a process that allows compliance with the European regulations while loosing as little information value as possible. Investigators must work jointly with the local Ethics Committees to harmonize the safeguarding of data privacy and security with the advances in clinical research which the BDA tools can offer us in future. Mechanisms must be developed to facilitate patient consent (or withdrawal of consent) to the use of the information generated during healthcare, conveniently harmonized for clinical research.

Another important issue is data property. In this case we are not dealing with privacy or security questions but with the intrinsic value of the information (even regarded as a commercial item). Does the patient (as the source of the information) or the clinician (as the subject entering the information in the system) have the right to request economical compensation or to prohibit the use of the data for clinical research purposes or any other use? Although the current reality is that the data are often in the hands of the companies that develop the medical software, and which seek to exploit such information for their own commercial interests,34 this issue is the subject of great controversy, and it would be desirable to establish legal norms seeking to resolve the problems we presently face in this field.

Big Data Analysis and Machine Learning tools offer a great opportunity for improving the strategic management of our Units, the handling of concrete clinical cases, and clinical research. However, in order to take advantage of this new methodology, we need to evolve and incorporate new human resources (specialized staff with clinical knowledge and training in AI) and technological assets. Furthermore, we must be able to harmonize the requirements referred to patient data privacy and security with the possibility of using large clinical databases efficiently.

We are convinced that an international and multidisciplinary team working collectively to draw knowledge from large bodies of clinical data, as explained in the course of this article, may bring a revolution to the modern practice of critical patient care.

The authors have received no financial support for the preparation of this article.

The authors declare that they have no conflicts of interest in relation to the contents of this article.