A percentage of people report prolonged and recurrent symptoms, for weeks or months, after the first episode of COVID-19. Persistent COVID-19 has not yet been precisely defined. It seems clear that it is a disease that affects a large number of people, generating a huge health and social impact in the pandemic [1]. In this context, patient groups, the international scientific community, as well as public health institutions and authorities are making a great effort to improve knowledge, clinical care, and social benefits [2,3]. Numerous Spanish scientific societies alongside patient groups have drawn up a Clinical Guide in which long COVID-19 is defined [2] as follows:

Determining the incidence of long COVID-19 is complicated due to the absence of specific surveillance and a variety of definitions. Another difficulty is that studies are performed in selected groups of patients, which does not allow estimating the true incidence in the population. The UK National Institute of Statistics estimated that 1 in 5 people with COVID-19 had symptoms beyond 5 weeks, and 1 in 10 people had symptoms beyond 12 weeks [4]. In a population-based survey in the United States, the percentage of people with persistent symptoms was 30% at 30 days, 25% at 60 days, and 15% at 90 days [5].

In the UK National Institute of Statistics study [6], a higher incidence was observed in women (23.6%, 95% CI 22.2-25.0) versus men (20.7%, 95% CI 19.3-22.1) and in the middle age. In the US survey, the only factor associated with persistent COVID-19, measured by the number of symptoms, was the initial severity [7].

The pathophysiological basis remains unknown, and several theories are put forward: the persistence of the virus in reservoirs, such as the small intestinal epithelium, where it would remain active [8,9]; the presence of an aberrant immune response [10]; the damage produced by the effect of autoantibodies against immunomodulatory proteins [11]; or the hyperactivation of coagulation and platelets [12]. The symptoms and signs are extremely numerous and varied (ie, systemic, neurological, psychiatric, and cardiovascular) [13-19].

It is important to know the people who experience long COVID-19 for a better characterization [20], since it is a characteristically intermittent disease with very varied symptoms, which also involves people who at the time did not have access to a diagnostic test and those affected who felt stigmatized and ignored [21].

A web-based survey by the Spanish Society of General and Family Physicians in 1834 participants reported a variety of more than 200 symptoms. These symptoms included fatigue and general malaise in more than 95% of the patients. Headache, low mood, and muscle aches were observed in more than 80% of them. Dyspnea, joint, chest, and back pain, as well as lack of concentration were detected in more than 75% of them. More than 70% found it difficult to attend their daily duties, and more than 30% reported difficulties even with personal hygiene. Although 52% of the cases were not confirmed by laboratory diagnostic tests, the authors noted that there were no significant differences between the groups with or without diagnostic confirmations [19].

COVID-19 caused serious problems to the health system, collapsing it, and depleting the health resources available [20-22]. Fortunately, the development and enforcement of specific public health policies, such as vaccination, mask wearing, and social distancing, among others, has reduced the prevalence and complications associated with COVID-19 in the acute phase [23]. However, the aftermath of the global pandemic requires an efficient approach to manage patients with long COVID-19. It seems a great opportunity to leverage on innovative digital health solutions to provide exhausted health care systems with the most cost-effective and efficient tools available to support the clinical management of the population experiencing long COVID-19 [24,25].

We can find more than 250 health-labeled apps available on both Google Play and Apple Store, but they are basic products that offer neither the technology nor the advanced approach and services offered to patients with long COVID-19. Adhera Health Fatigue Digital Program for long COVID-19 is built on the principles of patient centricity, and it is guided by the principles of participatory research to promote a meaningful partnership between patients and health care professionals [26]. Furthermore, data mining and artificial intelligence (AI) have been greatly applied lately to health care areas [27]. Data mining is a combination of statistical analysis, algorithms, AI, and database management, with the purpose of extracting intelligence. AI can be defined as the field devoted to build artificial creatures. It is the science and engineering of making intelligent machines, especially intelligent computer programs. In the last years, there has been a growing interest in the application of AI and data mining techniques to clinical data. MEDLINE has seen a sharp 10-fold increase in the number of papers having the term ‘data mining’ in their title [28]. This AI-driven research is designed toward the improvement of the Adhera Precision Digital Companion platform toward the provision of advanced personalization technologies for adaptive self-management, which is a progress beyond the state of the art [29,30].

The objective of this protocol is the prospective collection of psychometric and biometric data for training algorithms and prediction models to complement the SENSING-AI cohort. Likewise, the final aim of the project is the creation of a digital health solution based on AI and prediction models for a better clinical management of patients with long COVID-19 and to improve self-management of this condition.

This study was approved by the research ethics committee of Primary Care Research Institute Jordi Gol (in Barcelona, Spain) (Código CEIm: 22/010-PCV) and Virgen Macarena University Hospital (Seville, Spain) (1894-N-21). All patients will receive a patient information sheet and will sign an informed consent.

This study is registered in Clinicaltrials.gov website (NCT05204615). Data obtained from patients will be pseudo-anonymized by the clinical partner. Only a code, based on an alphanumeric number completely unlinked to any direct patient data, will be included as an identifier of the subjects in the study.

A prediction algorithm based on the nearest neighbor classification method will be used. This method is an instance-based algorithm supervised by machine learning. These algorithms will process the data flows in which the input is presented as a sequence of elements. Therefore, it will allow for searching in the closest observations. This algorithm cannot provide human interpretable models; processing procedures will be applied to make them explainable, based on feature classification. Therefore, this model has the aim of predicting whether the user is having a complication, based on clinical, biometric, and psychometric data.

By the data obtained in the prospective study, an adaptive adjustment of the sampling frequency of the ecological momentary assessments will be made. The objective is to develop a model to predict the most appropriate time to activate the validated questionnaire for the patient. This model will be developed using machine learning algorithms, mainly based on decision trees. These are flowchart-like structures in which each node represents a value in an entity, each branch represents the value, and each leaf represents a class or decision label after calculating all attributes. The model will be measured based on error rates and confusion tables, which will allow measuring accuracy, precision, F1 score, sensitivity, specificity, receiver operating characteristic curve, and area below the hamstring curve.

Another model will be developed to adapt the questionnaires to each patient. It will be focused on the user’s history and biometric measurements; the model will decide the number of questions needed for that patient. To develop this model, machine Learning models will be trained, mainly based on artificial neural networks. These are based on a collection of connected units or nodes called artificial neurons, which freely model the neurons in a biological brain. An artificial neuron receives a signal, processes it, and then signals to neurons connected to it. For each question, the input data will be the entities, and the output data will be a predicted answer to the question. If the question is easily predicted, the question will be removed from the test. If not, it has to remain in the test and be answered by the patient. This type of algorithm is not human interpretable, and therefore, postprocessing models will be applied to make them explainable.