Results

In accordance with the established criteria, we chose lung cancer as the pathology to study for the development of our framework. We did a detailed study of the local problem existing in the HUPHM with regard to CDA in lung cancer, and thus, it provided us with the following findings:

• 1. The registration process of healthcare data in the EMR is separated from the process of data registration for the research and care improving purpose, thus duplicating the registration work, inducing differences in interpretation and errors.
• 2. Beyond the EMR, there is no structured, unique and updated registry that contains a unique reference for each patient belonging to the oncology service for purposes of care improvement or research. There is a database of new patients (first consultations) of manual completion, fragmented into different databases over time, and not always updated in a timely manner. Nor is there a structured longitudinal register with the clinical follow-up information of the patients.
• 3. Activity data are available in those healthcare areas that are organized by the EMR (appointments in external consultations, inpatient admissions, day hospital, emergency department attendances, tests and interconsultations requested). In the areas in which the activity is coded (as in hospitalization through the mandatory minimum data set defined by the Spanish Ministry of Health, called CMBD), there are also clinical data associated with that activity. In the rest of the areas, only clinical data are available in the EMR (usually in free text), and in many cases, it is difficult to get activity data (e.g., cancer patients seen in the emergency department can be counted only through the interconsultations from ED to the Medical Oncology Department that are registered in the EMR).
• 4. The unstructured records of the EMR (in free text) require very intensive manual work to obtain the variables of interest and thus feed the RDBs.
• 5. There is no standard coding of structured values, nor a clear, written and shared definition by all users who enter data into the databases.

This affects the quality of the data and the possibilities to interoperate (integrate) with other records.

• 6. RDBs are created with specific objectives and therefore with specific data models for their purposes. However, they contain general patient data and specific variables that may be duplicated (and thus being inconsistent) between different RDBs. These RDBs do not have a unified reference that brings together a minimum set of oncological data (demographic and clinical).
• 7. We identify two main types of RDBs:
  • - RDBs created ad hoc for a specific research project. They are typically created after the definition of the project (hypothesis definition leads to necessary variables). They should contain a specific set of variables that were not available in other databases, but this is not always the case.
  • - RDBs contain care records that group patients with specific criteria for selection (patient cohorts), but without a closed definition a priori of the information that will be required for further research. They are created by the impossibility of locating these patients in the EMR by other means other than registering all the cases that meet criteria in a specific database. Sometimes, they are made with a core purpose, but can be used for further research (secondary use of information), serving as a list of patients to study (subsequently obtaining informed consent for this and also new variables once the research project has been fully determined).
  • - In both cases, the absence of common definitions of data and standardized values, of clear responsibilities regarding the creation and maintenance of subsequent data, together with a non-technical design, makes the data quality low in most cases. This requires additional efforts preparing data when it is decided to finally conduct an investigation.
• 8. In most cases, having a sufficiently structured patient database (a case register) would avoid the need to prepare different care records by pathology. When an investigation had to be conducted, a case registry query would be done and those patients that met criteria would be selected. From there, new specific information fields could be added for the desired investigation.
• 9. Although the function of the case registry could be assumed by the hospital’s EMR system, the reality is that the existing system (in the HUPHM and in many other Spanish hospitals, being also very similar to other national and international solutions) does not have the facilities for the autonomous management of the variables to be collected and for the exploitation of data that this unified case register would require. Although it can be proposed as a future possibility, it is necessary to propose a short-term alternative that allows progress in this area.
• 10. Although there are a large number of form templates in our EMR (846), up to 60% of them have been completed less than ten times. It is significant how variations are observed in different historical periods (with different management teams governing the hospital). Analysing the volume of records made perform template, it was observed how only four forms accumulated 73% of the total structured records of the EMR. These data suggest the low effective volume of clinical data that are structured in the EMR by the users themselves, beyond the data that is generated in a structured way in an automatic or semi-automatic way and of those that are captured as an essential part of a process of established work. In order to obtain better results in the structuring of the data by the users in the EMR, it is necessary that the users are willing to do so, either because they obtain a return of the additional effort that it entails against an unstructured registry, or because the structuring of data is an essential part of a work process established and enforced by the management team. In any case, the first option will always be the desirable one, since the second one can lead to situations of discontent and even burnout (Friedberg et al., 2013).
• 11. Regarding the volume of data, we observe that the number of patients of a specific pathology such as lung cancer, in a single centre, may not have sufficient volume to perform CDA using ML techniques, especially in case of selecting specific subpopulations. However, the large number of related entities (consultations, treatments, admissions, emergencies, etc.) and of variables that describe each of them over time (date of realization, performer, active principle, dose, discharge date, diagnosis date, etc.) does allow to have a high dimensionality and therefore a high volume of data for the application of advanced analysis techniques. In any case, the integration of information from different centres is essential for these technologies to offer their full potential.
• 12. Regarding the variety of data, another characteristic usually evaluated in the field of Big Data and DS, we observe that we mainly deal with clinical data collected as text or image. However, we can integrate other sources of information. Recently, the development of personalized medicine in oncology has promoted the incorporation of molecular

• (genetic) information, with the near perspective of exponential growth, by having mechanisms of massive sequencing and analysis of proteins or microbiome that could be of interest to establish a unique profile of the patient at each time. At the same time, the integration in the future of other data of the individual, derived from its location in a population context, such as cultural, socioeconomic and demographic, epidemiological and public health data, habits and behaviours or values and preferences (the so-called social determinants of health or SDOH) is foreseeable.
• 13. Finally, in relation to the speed of data generation or the generation of a response to them (whether generated knowledge, a prediction or an action), we observe a growing need. However, the bulk of the data identified in our work does not have special processing speed requirements.