Background. In type 2 diabetes mellitus (T2DM), cardiovascular mortality is twice as high in men and four times higher in women. The American Heart Association (AHA) considers diabetes an equivalent of cardiovascular risk, emphasizing that patients with T2DM without a previous history of myocardial infarction have a similar risk of coronary artery disease and cardiovascular events to non-diabetic individuals who have already experienced an event. Objective. This study aims to assess the ability of an artificial intelligence (AI)-based algorithm in predicting the risk of cardio-cerebrovascular complications in patients with T2DM. Subjects and methods. The digital medical records of 532 patients from the Diabetology Unit U.O.S.D. at AULSS 6 Euganea were analyzed. Clinical data for each patient were managed using the software MetaClinic, which was enhanced with the ‘AI Prediction Module’. This module calculates the probability of organ damage in six key organs commonly affected by chronic diabetes complications. The study specifically focused on analyzing the datasets related to the “heart” and “cerebral vessels.” For patients for whom the AI algorithm predicted a “very high” or “low” risk of developing heart disease, clinical and instrumental data related to their cardiac history were also collected, in addition to the (AI)-generated risk predictions. The data were processed using Cohen’s K statistical method to analyze the agreement between (AI)-predictive data and traditional clinical-instrumental diagnostics (ECG, carotid ultrasound with color Doppler, medical history). Results. The concordance analysis between the (AI)-based risk prediction and clinical-instrumental data was initially conducted on two groups: patients with a “very high” risk of heart disease and patients with a “low” risk of heart disease. In the first group, Cohen’s K coefficient for heart disease was K=0,00, indicating no agreement between the (AI)-prediction and clinical-instrumental data, whereas, for cerebrovascular disease, the coefficient was K=0,89, demonstrating a high level of agreement between the predictive algorithm and measurable data. A K=0,00 was again observed for heart disease in the second group, confirming the lack of agreement between (AI)-prediction and traditional diagnostics. In contrast, a value of K=0,83 was afresh recorded for cerebrovascular disease, confirming a high level of agreement between evaluators. Since the analysis of the two patient samples was significantly affected by presence of null values in contingency tables, it was deemed appropriate to conduct a combined analysis that included both patient samples as a single group. In the aggregated evaluation, Cohen’s K coefficient for heart disease reached a value of K=0,09, still indicating a very low level of agreement; in contrast, for cerebrovascular disease, the coefficient remained high at K=0,85, reiteratively confirming strong concordance between the predictive algorithm and clinical-instrumental diagnostics. Conclusions. The results highlight a marked difference in the algorithm’s ability concerning the two examined diabetes-related complications. In comparison, while (AI)-prediction for cerebrovascular disease showed a high level of agreement with traditional diagnostic methods, the algorithm’s predictive ability proved insufficient in assessing the risk of heart disease. It emphasizes refining the model to enhance its precision in stratifying cardio-cerebrovascular risk. However, the current examination shows several limitations that were comprehensively explored in the discussion.
Background. Nel diabete mellito di tipo 2 (DMT2), la mortalità cardiovascolare è doppia negli uomini e quadruplicata nelle donne. L’American Heart Association (AHA) considera il diabete un equivalente di rischio cardiovascolare, sottolineando come i pazienti affetti senza pregressa storia di infarto del miocardio presentino un rischio di coronaropatia ed eventi cardiovascolari simile a quello di soggetti non diabetici che hanno già subito un evento. Scopo dello studio. L’obiettivo di questo studio è valutare la capacità di un algoritmo, basato sull’intelligenza artificiale (IA), nel predire il rischio di complicanze cardio-cerebrovascolari in una coorte di pazienti affetti da DMT2. Soggetti e metodi. Sono state analizzate le cartelle cliniche digitali di 532 pazienti seguiti presso l’U.O.S.D. di Diabetologia dell’AULSS 6 Euganea. Il software MetaClinic, attualmente in uso per la gestione dei dati clinici, è stato implementato utilizzando il ‘Modulo di Predizione IA’, dedicato al calcolo della probabilità di insorgenza di danno d’organo nei sei principali organi interessati da complicanze croniche del diabete mellito. Lo studio si è poi orientato sull’analisi delle schede “cuore” e “vasi cerebrali”. Oltre alle predizioni fornite dall’IA, sono stati raccolti dati clinici e strumentali relativi alla storia cardiologica dei soli pazienti per cui la predizione di rischio fornita dall’algoritmo per il rischio di sviluppare cardiopatia fosse risultata “molto alta” oppure “bassa”. I dati sono stati elaborati con metodo statistico K di Cohen, con la finalità di fornire un’analisi di concordanza tra dati predittivi di IA e diagnostica tradizionale clinico-strumentale (ECG, ecocolordoppler TSA, storia clinica). Risultati. L’analisi di concordanza è stata condotta inizialmente su due gruppi isolati: pazienti con rischio “molto alto” e pazienti con rischio “basso” di cardiopatia. Nel primo gruppo, il coefficiente K di Cohen per cardiopatia è risultato pari a K=0,00, indicando un’assenza di concordanza tra predizione dell’IA e dati clinico-strumentali, mentre per il rischio di vasculopatia cerebrale è risultato pari a K=0,89, evidenziando un’elevata concordanza tra l’algoritmo predittivo e i dati misurabili. Nel secondo gruppo si è riscontrato un K=0,00 per il rischio di cardiopatia, confermando la mancanza di accordo tra predizione dell’IA e diagnostica tradizionale, mentre per il rischio di vasculopatia cerebrale si è ottenuto un valore di K=0,83, che ha confermato un elevato livello di concordanza tra i valutatori. Poiché l’analisi dei due campioni di pazienti risultava notevolmente influenzata dalla presenza di valori nulli nelle tabelle di contingenza, si è ritenuto opportuno procedere con un’analisi complessiva che includesse entrambi i campioni in un unico gruppo. Nella valutazione aggregata, il coefficiente K di Cohen per la cardiopatia ha raggiunto un valore di K=0,09, indicando un livello di concordanza ancora molto basso, mentre per la valutazione del rischio di vasculopatia cerebrale è rimasto elevato con un valore pari a K=0,85. Conclusioni. I risultati evidenziano una marcata differenza nella capacità predittiva dell’algoritmo in relazione alle due complicanze esaminate: per la vasculopatia cerebrale si è riscontrata un’elevata concordanza tra la predizione dell’IA e i metodi diagnostici tradizionali, mentre la capacità predittiva dell’algoritmo nella valutazione del rischio di cardiopatia si è rivelata insufficiente. Ciò suggerirebbe la necessità di ottimizzare il modello al fine di migliorarne l’accuratezza nella stratificazione del rischio cardio-cerebrovascolare, tuttavia lo studio presenta alcuni limiti che sono stati approfonditi nella discussione.
L’INTELLIGENZA ARTIFICIALE COME STRUMENTO PER PREDIRE IL RISCHIO CARDIO-CEREBROVASCOLARE NEL PAZIENTE AFFETTO DA DMT2: UN APPROCCIO INNOVATIVO BASATO SU ESAMI BIOUMORALI. ANALISI DI CONCORDANZA CON DIAGNOSTICA CLINICO-STRUMENTALE.
CELSAN, CHIARA CELESTE
2024/2025
Abstract
Background. In type 2 diabetes mellitus (T2DM), cardiovascular mortality is twice as high in men and four times higher in women. The American Heart Association (AHA) considers diabetes an equivalent of cardiovascular risk, emphasizing that patients with T2DM without a previous history of myocardial infarction have a similar risk of coronary artery disease and cardiovascular events to non-diabetic individuals who have already experienced an event. Objective. This study aims to assess the ability of an artificial intelligence (AI)-based algorithm in predicting the risk of cardio-cerebrovascular complications in patients with T2DM. Subjects and methods. The digital medical records of 532 patients from the Diabetology Unit U.O.S.D. at AULSS 6 Euganea were analyzed. Clinical data for each patient were managed using the software MetaClinic, which was enhanced with the ‘AI Prediction Module’. This module calculates the probability of organ damage in six key organs commonly affected by chronic diabetes complications. The study specifically focused on analyzing the datasets related to the “heart” and “cerebral vessels.” For patients for whom the AI algorithm predicted a “very high” or “low” risk of developing heart disease, clinical and instrumental data related to their cardiac history were also collected, in addition to the (AI)-generated risk predictions. The data were processed using Cohen’s K statistical method to analyze the agreement between (AI)-predictive data and traditional clinical-instrumental diagnostics (ECG, carotid ultrasound with color Doppler, medical history). Results. The concordance analysis between the (AI)-based risk prediction and clinical-instrumental data was initially conducted on two groups: patients with a “very high” risk of heart disease and patients with a “low” risk of heart disease. In the first group, Cohen’s K coefficient for heart disease was K=0,00, indicating no agreement between the (AI)-prediction and clinical-instrumental data, whereas, for cerebrovascular disease, the coefficient was K=0,89, demonstrating a high level of agreement between the predictive algorithm and measurable data. A K=0,00 was again observed for heart disease in the second group, confirming the lack of agreement between (AI)-prediction and traditional diagnostics. In contrast, a value of K=0,83 was afresh recorded for cerebrovascular disease, confirming a high level of agreement between evaluators. Since the analysis of the two patient samples was significantly affected by presence of null values in contingency tables, it was deemed appropriate to conduct a combined analysis that included both patient samples as a single group. In the aggregated evaluation, Cohen’s K coefficient for heart disease reached a value of K=0,09, still indicating a very low level of agreement; in contrast, for cerebrovascular disease, the coefficient remained high at K=0,85, reiteratively confirming strong concordance between the predictive algorithm and clinical-instrumental diagnostics. Conclusions. The results highlight a marked difference in the algorithm’s ability concerning the two examined diabetes-related complications. In comparison, while (AI)-prediction for cerebrovascular disease showed a high level of agreement with traditional diagnostic methods, the algorithm’s predictive ability proved insufficient in assessing the risk of heart disease. It emphasizes refining the model to enhance its precision in stratifying cardio-cerebrovascular risk. However, the current examination shows several limitations that were comprehensively explored in the discussion.File | Dimensione | Formato | |
---|---|---|---|
Celsan_ChiaraCeleste.pdf
accesso riservato
Dimensione
5.24 MB
Formato
Adobe PDF
|
5.24 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/82874