Introduction. The Covid-19 pandemic had a substantial impact on older adults and frail individuals, supporting the importance of early identification of risk factors for adverse outcomes. Although these population have been widely investigated using traditional statistical methods, the application of Artificial Intelligence (AI) in these settings remains limited. This thesis consists of two distinct epidemiological studies conducted on older adults living in LTC (Study 1) and patients hospitalized for acute Covid-19 (Study 2), whose statistical analyses were performed using AI models. Study 1. Objective: To investigate, using AI, the factors associated with adverse outcomes (Covid-19 infection, oxygen therapy, and death) and with a high post-vaccination antibody response among older adults vaccinated against SARS-CoV-2 and living in LTC. Methods: Multicenter longitudinal study conducted between March 2021 and June 2022 among LTC residents vaccinated against SARS-CoV-2. Results: The mean age of the 3276 participants was 83.4 years, and 71.0% were women. Over 12 months of follow-up, 9.3% developed Covid-19, 1.4% required oxygen therapy, and 15.6% died. The best predictive model for incident Covid -19 infection was Random Forest (RF). The feature most strongly associated with a reduced likelihood of infection was a previous SARS-CoV-2 infection; having received two vaccine doses also reduced the risk of incident infection. In the subsample with antibody measurements, XGBoost and RF showed that, both at 2 and 6 months, a high immune response was associated with a previous SARS-CoV-2 infection, immunization with the Moderna vaccine (vs Pfizer/BioNTech), and non-smoking status (likely due to survival bias). Reliable predictive models for oxygen therapy and death could not be developed because of low performance metrics. Study 2. Objective: To evaluate, using AI, the association between specific genetic loci and 6-month mortality (main outcome) and clinical severity of acute disease and in-hospital delirium (secondary outcomes) among adults hospitalized for Covid-19. Methods: Single-center longitudinal study conducted in the Internal Medicine Unit of the University Hospital of Ferrara (Italy) among adults hospitalized for acute Covid-19 between March 1, 2020, and March 1, 2021. Results: The mean age of the 363 participants was 71.2 years, and 47.6% were women. During hospitalization, 8.3% developed delirium, 23.3% required escalation of health care, and 6-month mortality was 28.0%. According to the Random Survival Forest (RSF) - the best predictive model for survival analysis, the two genetic loci most strongly associated with mortality were ACE2 and PPARGC1A; however, their predictive contribution was weaker than that of age, number of chronic medications, home setting prior to admission (vs healthcare facility), functional dependence, arterial hypertension, and dementia. Feature analysis was not conducted for escalation of health care and delirium due to poor predictive performance of the models. Conclusions. Among older adults in LTC, previous Covid-19 infection was the main protective factor against incident infection within 12 months and was associated with a high antibody response. In patients hospitalized for acute SARS-CoV-2 infection, clinical factors remained stronger predictors of mortality than genetic ones, although ACE2 and PPARGC1A showed relevant associations. AI appears to be a promising tool in population studies, but a multidisciplinary approach is required to ensure the validity and clinical applicability of the findings.
Introduzione. La pandemia da Covid-19 ha avuto un impatto significativo sulle persone anziane e sui soggetti fragili, evidenziando l’importanza di identificare precocemente i fattori di rischio di esiti avversi. Sebbene tali popolazioni siano state ampiamente studiate con metodi di statistica tradizionale, l’uso dell’Intelligenza Artificiale (AI) in questi ambiti rimane limitato. La tesi include due studi epidemiologici distinti condotti su residenti in Lungodegenza Post-Acuzie (LPA) (Studio 1) e ricoverati per malattia acuta da Covid-19 (Studio 2), la cui elaborazione statistica è stata volta tramite modelli di AI. Studio 1. Obiettivo: Indagare tramite AI i fattori associati ad outcome avversi (infezione da Covid-19, necessità di ossigenoterapia e decesso) e a una più alta risposta anticorpale a vaccino in anziani vaccinati con anti-SARS-CoV-2 residenti in regime di LPA. Metodi: Studio multicentrico longitudinale svolto fra marzo 2021 e giugno 2022 su residenti in LPA vaccinati con anti-SARS-CoV-2. Risultati: L’età media dei 3276 partecipanti è stata di 83.4 anni e il 71.0% era di sesso femminile. Nei 12 mesi di follow-up, il 9.3% dei pazienti ha sviluppato infezione da Covid-19, il 1.4% ha richiesto ossigeno, mentre il 15.6% è deceduto. Il modello predittivo migliore per infezione incidente da Covid-19 è risultato il Random Forest (RF). La feature maggiormente associata a tale outcome è stata la pregressa infezione da SARS-CoV-2, che ne riduceva la probabilità; inoltre, anche aver ricevuto due dosi di vaccino anti-SARS-CoV-2 riduceva il rischio di infezione incidente. Nel sottocampione con dosaggio anticorpale, i XGBoost e RF hanno mostrato che, sia a 2 che a 6 mesi, una risposta immunitaria alta era associata a una pregressa infezione da SARS-CoV-2, al vaccino Moderna (vs. Pfizer/BioNTech) e allo stato di non fumatore (verosimile survival bias). Non è stato possibile costruire modelli predittivi per ossigenoterapia e decesso per basse metriche di performance. Studio 2. Obiettivi: Valutare tramite AI la presenza di un’associazione fra specifici loci genetici e decesso entro 6 mesi (main outcome) e necessità di intensificazione delle cure e delirium intraospedaliero (secondary outcomes) in adulti ospedalizzati per infezione da Covid-19. Metodi: Studio monocentrico longitudinale svolto nella U.O. di Medicina Interna Universitaria dell’Azienda Ospedaliero-Universitaria Sant’Anna di Ferrara (Italia) su adulti ospedalizzati con malattia acuta da Covid-19 fra l’1 marzo 2020 e l’1 marzo 2021. Risultati: L’età media dei 363 partecipanti è stata di 71.2 anni e il 47.6% era di sesso femminile. L’8.3% dei pazienti ha avuto delirium durante il ricovero, il 23.3% ha richiesto un’intensificazione delle cure, mentre la mortalità a 6 mesi è stata del 28.0%. In base al Random Survival Forest (RSF), risultato il migliore modello predittivo nell’analisi di sopravvivenza, i due loci genetici maggiormente associati a decesso sono stati ACE2 e PPARGC1A; la forza di tali associazioni è risultata però minore di quella data da età, numero di farmaci, domicilio come setting pre-ricovero, dipendenza funzionale, ipertensione e demenza. Relativamente alla richiesta di intensificazione di cure e delirium intraricovero, l’analisi delle feature non è stata svolta, per basse metriche di performance. Conclusioni. Negli anziani in LPA, una precedente infezione da Covid-19 è risultato il principale fattore protettivo verso l’infezione incidente entro 12 mesi ed associato a migliore risposta anticorpale. Nei pazienti ricoverati per malattia acuta da SARS-CoV-2, i fattori clinici restano predittori di mortalità più forti rispetto a quelli genetici, sebbene ACE2 e PPARGC1A mostrino associazioni rilevanti. L’AI emerge come strumento promettente negli studi di popolazione, ma richiede un approccio multidisciplinare per garantire la validità e l’applicabilità dei risultati nella pratica clinica.
Applicazione di tecniche di analisi di Intelligenza Artificiale nello studio di stato di salute, epidemiologia ed esposizione per l’identificazione dei fattori di rischio individuali nel Covid-19 e nella risposta ai vaccini
REMELLI, Francesca
2026
Abstract
Introduction. The Covid-19 pandemic had a substantial impact on older adults and frail individuals, supporting the importance of early identification of risk factors for adverse outcomes. Although these population have been widely investigated using traditional statistical methods, the application of Artificial Intelligence (AI) in these settings remains limited. This thesis consists of two distinct epidemiological studies conducted on older adults living in LTC (Study 1) and patients hospitalized for acute Covid-19 (Study 2), whose statistical analyses were performed using AI models. Study 1. Objective: To investigate, using AI, the factors associated with adverse outcomes (Covid-19 infection, oxygen therapy, and death) and with a high post-vaccination antibody response among older adults vaccinated against SARS-CoV-2 and living in LTC. Methods: Multicenter longitudinal study conducted between March 2021 and June 2022 among LTC residents vaccinated against SARS-CoV-2. Results: The mean age of the 3276 participants was 83.4 years, and 71.0% were women. Over 12 months of follow-up, 9.3% developed Covid-19, 1.4% required oxygen therapy, and 15.6% died. The best predictive model for incident Covid -19 infection was Random Forest (RF). The feature most strongly associated with a reduced likelihood of infection was a previous SARS-CoV-2 infection; having received two vaccine doses also reduced the risk of incident infection. In the subsample with antibody measurements, XGBoost and RF showed that, both at 2 and 6 months, a high immune response was associated with a previous SARS-CoV-2 infection, immunization with the Moderna vaccine (vs Pfizer/BioNTech), and non-smoking status (likely due to survival bias). Reliable predictive models for oxygen therapy and death could not be developed because of low performance metrics. Study 2. Objective: To evaluate, using AI, the association between specific genetic loci and 6-month mortality (main outcome) and clinical severity of acute disease and in-hospital delirium (secondary outcomes) among adults hospitalized for Covid-19. Methods: Single-center longitudinal study conducted in the Internal Medicine Unit of the University Hospital of Ferrara (Italy) among adults hospitalized for acute Covid-19 between March 1, 2020, and March 1, 2021. Results: The mean age of the 363 participants was 71.2 years, and 47.6% were women. During hospitalization, 8.3% developed delirium, 23.3% required escalation of health care, and 6-month mortality was 28.0%. According to the Random Survival Forest (RSF) - the best predictive model for survival analysis, the two genetic loci most strongly associated with mortality were ACE2 and PPARGC1A; however, their predictive contribution was weaker than that of age, number of chronic medications, home setting prior to admission (vs healthcare facility), functional dependence, arterial hypertension, and dementia. Feature analysis was not conducted for escalation of health care and delirium due to poor predictive performance of the models. Conclusions. Among older adults in LTC, previous Covid-19 infection was the main protective factor against incident infection within 12 months and was associated with a high antibody response. In patients hospitalized for acute SARS-CoV-2 infection, clinical factors remained stronger predictors of mortality than genetic ones, although ACE2 and PPARGC1A showed relevant associations. AI appears to be a promising tool in population studies, but a multidisciplinary approach is required to ensure the validity and clinical applicability of the findings.I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


