Article type
Abstract
"Background
Timely syntheses of evidence are necessary for making informed decisions, particularly regarding high-priority urgent health problems.
Objectives
This study aims to evaluate the functionality and efficiency of Nested Knowledge, a semi-automated platform powered by artificial intelligence (AI ) for conducting SRs.
Methods
We used Nested Knowledge, a semi-automated web-based platform to conduct systematic reviews (CRD42023475592). This software is composed of two parts which work in tandem. AutoLit® streamlines the research process, facilitating efficient literature search, screening, data extraction, and critical appraisal. It also provides insightful data visualization, analysis, and knowledge dissemination. We initially used it to select studies assessing how rapid correction compares with slow correction in terms of mortality, neurological complications, and length of stay.
Results
Our search strategy retrieved 5010 records and finally included 18 studies. A PRISMA chart was automatically generated (Figure 1). The platform was very intuitive, and no special training was required. The automatic PICO highlighting (including own keywords), and the AI-prediction of the most relevant studies to our research question facilitated the screening by title and abstract. After training the model with 50 records (including 10 potentially eligible studies), we used dual independent screening. Once one reviewer screened the remaining 4960 records, the dual screening by robot screener took <1 minute and the discrepancy resolution of 292 records took ~3 hours. The cross-validation statistics of the Inclusion Probability Model reported a Recall of 0.65 and AUC of 0.98. The robot excluded no potentially eligible study. Smart tag recommendations using GPT4 automatically highlight full texts based on our configured tags. It found 68.4% of the original extracted variables when compared to the human extraction. The platform maintained a full audit record of our activities and organized our interventions and data of interest. The performance of the rest of the functionalities will be explored, and presented at the Colloquium.
Conclusions: We found Nested Knowledge to be highly efficient for selecting studies. "
Timely syntheses of evidence are necessary for making informed decisions, particularly regarding high-priority urgent health problems.
Objectives
This study aims to evaluate the functionality and efficiency of Nested Knowledge, a semi-automated platform powered by artificial intelligence (AI ) for conducting SRs.
Methods
We used Nested Knowledge, a semi-automated web-based platform to conduct systematic reviews (CRD42023475592). This software is composed of two parts which work in tandem. AutoLit® streamlines the research process, facilitating efficient literature search, screening, data extraction, and critical appraisal. It also provides insightful data visualization, analysis, and knowledge dissemination. We initially used it to select studies assessing how rapid correction compares with slow correction in terms of mortality, neurological complications, and length of stay.
Results
Our search strategy retrieved 5010 records and finally included 18 studies. A PRISMA chart was automatically generated (Figure 1). The platform was very intuitive, and no special training was required. The automatic PICO highlighting (including own keywords), and the AI-prediction of the most relevant studies to our research question facilitated the screening by title and abstract. After training the model with 50 records (including 10 potentially eligible studies), we used dual independent screening. Once one reviewer screened the remaining 4960 records, the dual screening by robot screener took <1 minute and the discrepancy resolution of 292 records took ~3 hours. The cross-validation statistics of the Inclusion Probability Model reported a Recall of 0.65 and AUC of 0.98. The robot excluded no potentially eligible study. Smart tag recommendations using GPT4 automatically highlight full texts based on our configured tags. It found 68.4% of the original extracted variables when compared to the human extraction. The platform maintained a full audit record of our activities and organized our interventions and data of interest. The performance of the rest of the functionalities will be explored, and presented at the Colloquium.
Conclusions: We found Nested Knowledge to be highly efficient for selecting studies. "