Background: The fight against hospital-acquired infections (HAIs) and the related antimicrobial resistance is among the key priorities of the World Health Organization. According to the CDC definition, a surgical site infection(SSI) is an infection that occurs in the part of the body where the surgery took place within 30 days after surgery or 90 days if there is a prosthetic implant. SSIs account for 20% of all HAIs and are a common postoperative complication. In Europe, 1% to 10% of patients develop a SSI among those who undergo surgical procedures, depending upon the operative site and wound classification. Problem: SSIs cause a substantial burden worldwide in terms of patient morbidity, mortality, and additional costs. A considerable proportion of SSI could be avoided through the implementation of appropriate surveillance and control programs. Surveillance systems for intra-hospital SSIs usually underestimate SSI incidence and one of the main reasons is that laboratory data are not sufficient for SSI diagnosis. According to the SSI definition, clinical reporting plays an important role in identifying SSIs and this is not commonly integrated into existing automatic surveillance systems, even if it could increase their sensitivity. Around 50% of SSIs occur after discharge, those can be severe and hard to predict using common risk indices. Solution: The first step for SSI prevention is to develop an automatic surveillance system to identify intra-hospital infections, for this task (task 1) we are training a Natural Language Processing (NLP) algorithm using unstructured text data from 22,625 hospital discharge letters of the University Hospital of Pisa, Italy, from 2020 and 2021. The subsequent task (task 2) is to develop a risk-stratification algorithm for postdischarge SSI, based on the hospital discharge letters. A supervised learning approach can be used for both classification tasks, but in this process labeling can be a challenging step. For task 1, around 40 Public Health residents have been recruited and instructed to label the post-discharge letters in three classes: SSI, No SSI, or SSI in a precedent hospitalization. Data on post-discharge SSI should be collected from regional or national SSI surveillance programs, to obtain a ground truth necessary for task 2. This AI-based surveillance system could become a valuable tool to address the lack of adequate SSI surveillance in our hospital, addressing the underestimation issue. The algorithm could then be fine-tuned with data collected from other hospitals and tested in other settings, using a federated learning approach. In addition to surveillance, the identification of patients at high risk of postdischarge SSI can guide patient-tailored preventive strategies.
Hospital-acquired infections surveillance and prevention: using Natural Language Processing to analyze unstructured text of hospital discharge letters for surgical site infections identification and risk-stratification
De Angelis, Luigi;Baglivo, Francesco;Arzilli, Guglielmo;Calamita, Leonardo;Ferragina, Paolo;Privitera, Gaetano Pierpaolo;Rizzo, Caterina
2023-01-01
Abstract
Background: The fight against hospital-acquired infections (HAIs) and the related antimicrobial resistance is among the key priorities of the World Health Organization. According to the CDC definition, a surgical site infection(SSI) is an infection that occurs in the part of the body where the surgery took place within 30 days after surgery or 90 days if there is a prosthetic implant. SSIs account for 20% of all HAIs and are a common postoperative complication. In Europe, 1% to 10% of patients develop a SSI among those who undergo surgical procedures, depending upon the operative site and wound classification. Problem: SSIs cause a substantial burden worldwide in terms of patient morbidity, mortality, and additional costs. A considerable proportion of SSI could be avoided through the implementation of appropriate surveillance and control programs. Surveillance systems for intra-hospital SSIs usually underestimate SSI incidence and one of the main reasons is that laboratory data are not sufficient for SSI diagnosis. According to the SSI definition, clinical reporting plays an important role in identifying SSIs and this is not commonly integrated into existing automatic surveillance systems, even if it could increase their sensitivity. Around 50% of SSIs occur after discharge, those can be severe and hard to predict using common risk indices. Solution: The first step for SSI prevention is to develop an automatic surveillance system to identify intra-hospital infections, for this task (task 1) we are training a Natural Language Processing (NLP) algorithm using unstructured text data from 22,625 hospital discharge letters of the University Hospital of Pisa, Italy, from 2020 and 2021. The subsequent task (task 2) is to develop a risk-stratification algorithm for postdischarge SSI, based on the hospital discharge letters. A supervised learning approach can be used for both classification tasks, but in this process labeling can be a challenging step. For task 1, around 40 Public Health residents have been recruited and instructed to label the post-discharge letters in three classes: SSI, No SSI, or SSI in a precedent hospitalization. Data on post-discharge SSI should be collected from regional or national SSI surveillance programs, to obtain a ground truth necessary for task 2. This AI-based surveillance system could become a valuable tool to address the lack of adequate SSI surveillance in our hospital, addressing the underestimation issue. The algorithm could then be fine-tuned with data collected from other hospitals and tested in other settings, using a federated learning approach. In addition to surveillance, the identification of patients at high risk of postdischarge SSI can guide patient-tailored preventive strategies.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
