ABSTRACT
Since February 2020, the Italian National Healthcare System had to mitigate the possibility of Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) transmission to vulnerable patients. Healthcare professionals rapidly reviewed their workflow to maintain a safe and high standard treatment, but weak scientific evidences and organizational limits resulted in the adoption of heterogeneous measures. Adherence to screening protocols and follow-up programs pregnant women and oncological patients has not been always guaranteed: this scenario could evolve in an enormous number of medico-legal actions. This context, showing the weakness of the Italian law No. 24/2017, imposes an urgent reorganization of the legal framework to homogenize the judgements to “protect” healthcare professionals involved in this epochal emergency.
ABSTRACT
Background Primary care physicians have a crucial role in determining the appropriate healthcare setting for their confirmed or suspect COVID-19 patients. Machine learning provides science-based tools that can be used for clinical decision-making which have already been applied to the fight against Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) both in the therapeutic and the prevention area. The aim of this study was to develop a machine learning-based tool to support primary care physicians to decide between home monitoring and hospitalization for their patients before diagnostic test results are available. Methods A retrospective cohort study with data from a hospital setting was performed. Patients' medical history and clinical, laboratory and radiological findings were collected and the dataset was used to train a predictive model for COVID-19 severity. The patients were divided between confirmed and suspect cases on the basis of the positivity of the nasopharyngeal RT-PCR test results. A splitting algorithm was recursively used to choose the predictor. A decision tree was built. Results A total of 198 subjects were enrolled for the study. Out of them, 28 cases were classified as mild disease, 62 as moderate disease, 64 as severe disease, and 44 as critical disease, according to WHO guidelines. The G2 value was used to determine the contribution of each obtained value to build the decision tree. The tree was, therefore, built choosing values that maximized G2 and LogWorth. SpO2 (cut point = 92%) was chosen for the optimal first split. The correspondence between inputs and outcomes was validated. Conclusions Our tool provides accurate clinical severity prediction for both confirmed and suspect COVID-19 patients. We, therefore, propose its implementation in the everyday life challenges of primary care physicians to support their clinical decision-making in providing appropriate and timely care for their patients. Key messages Primary care physicians have a crucial role in determining the appropriate healthcare setting for their confirmed or suspect COVID-19 patients. We propose a tool that provides an accurate clinical severity prediction for both confirmed and suspect COVID-19 patients to help choosing the appropriate healthcare setting for them.
ABSTRACT
OBJECTIVE: To develop a deep learning-based decision tree for the primary care setting, to stratify adult patients with confirmed and unconfirmed coronavirus disease 2019 (COVID-19), and to predict the need for hospitalization or home monitoring. PATIENTS AND METHODS: We performed a retrospective cohort study on data from patients admitted to a COVID hospital in Rome, Italy, between 5 March 2020 and 5 June 2020. A confirmed case was defined as a patient with a positive nasopharyngeal RT-PCR test result, while an unconfirmed case had negative results on repeated swabs. Patients' medical history and clinical, laboratory and radiological findings were collected, and the dataset was used to train a predictive model for COVID-19 severity. RESULTS: Data of 198 patients were included in the study. Twenty-eight (14.14%) had mild disease, 62 (31.31%) had moderate disease, 64 (32.32%) had severe disease, and 44 (22.22%) had critical disease. The G2 value assessed the contribution of each collected value to decision tree building. On this basis, SpO2 (%) with a cut point at 92 was chosen for the optimal first split. Therefore, the decision tree was built using values maximizing G2 and LogWorth. After the tree was built, the correspondence between inputs and outcomes was validated. CONCLUSIONS: We developed a machine learning-based tool that is easy to understand and apply. It provides good discrimination in stratifying confirmed and unconfirmed COVID-19 patients with different prognoses in every context. Our tool might allow general practitioners visiting patients at home to decide whether the patient needs to be hospitalized.