1.
A fusion decision system to identify and grade malnutrition in cancer patients: Machine learning reveals feasible workflow from representative real-world data.
Yin, L, Song, C, Cui, J, Lin, X, Li, N, Fan, Y, Zhang, L, Liu, J, Chong, F, Wang, C, et al
Clinical nutrition (Edinburgh, Scotland). 2021;(8):4958-4970
Abstract
BACKGROUND AND AIMS Most nutritional assessment tools are based on pre-defined questionnaires or consensus guidelines. However, it has been postulated that population data can be used directly to develop a solution for assessing malnutrition. This study established a machine learning (ML)-based, individualized decision system to identify and grade malnutrition using large-scale data from cancer patients. METHODS This was an observational, nationwide, multicenter cohort study that included 14134 cancer patients from five institutions in four different geographic regions of China. Multi-stage K-means clustering was performed to isolate and grade malnutrition based on 17 core nutritional features. The effectiveness of the identified clusters for reflecting clinical characteristics, nutritional status and patient outcomes was comprehensively evaluated. The study population was randomly split for model derivation and validation. Multiple ML algorithms were developed, validated and compared to screen for optimal models to implement the cluster prediction. RESULTS A well-nourished cluster (n = 8193, 58.0%) and a malnourished cluster with three phenotype-specific severity levels (mild = 2195, 15.5%; moderate = 2491, 17.6%; severe = 1255, 8.9%) were identified. The clusters showed moderate agreement with the Patient-Generated Subjective Global Assessment and the Global Leadership Initiative on Malnutrition criteria. The severity of malnutrition was negatively associated with the nutritional status, physical status, quality of life, and short-term outcomes, and was monotonically correlated with reduced overall survival. A multinomial logistic regression was found to be the optimal ML algorithm, and models built based on this algorithm showed almost perfect performance to predict the clusters in the validation data. CONCLUSIONS This study developed a fusion decision system that can be used to facilitate the identification and severity grading of malnutrition in patients with cancer. Moreover, the study workflow is flexible, and might provide a generalizable solution for the artificial intelligence-based assessment of malnutrition in a wider variety of scenarios.
2.
Quantitative contrast-enhanced harmonic EUS in differential diagnosis of focal pancreatic masses (with videos).
Săftoiu, A, Vilmann, P, Dietrich, CF, Iglesias-Garcia, J, Hocke, M, Seicean, A, Ignee, A, Hassan, H, Streba, CT, Ioncică, AM, et al
Gastrointestinal endoscopy. 2015;(1):59-69
Abstract
BACKGROUND The role of EUS with contrast agents can be expanded through the use of time-intensity curve (TIC) analysis and computer-aided interpretation. OBJECTIVE To validate the use of parameters derived from TIC analysis in an artificial neural network (ANN) classification model designed to diagnose pancreatic carcinoma (PC) and chronic pancreatitis (CP). SETTING Prospective, multicenter, observational trial-endoscopy units from Romania, Denmark, Germany, and Spain. PATIENTS A total of 167 consecutive patients with PC or CP. INTERVENTIONS Contrast-enhanced harmonic EUS (CEH-EUS) and EUS-guided FNA (EUS-FNA), TIC analysis, and ANN processing. MAIN OUTCOME MEASUREMENTS Sensitivity, specificity, positive and negative predictive values (PPV, NPV) for EUS-FNA, CEH-EUS, and the ANN. RESULTS After excluding all of the recordings that did not meet the technical and procedural criteria, 112 cases of PC and 55 cases of CP were included. EUS-FNA was performed in 129 patients, and the diagnosis was confirmed by surgery (n = 15) or follow-up (n = 23) in the remaining cases. Its sensitivity and specificity were 84.82% and 100%, respectively, whereas the PPV and NPV were 100% and 76.63%, respectively. The sensitivity of real-time quantitative assessment of CEH-EUS was 87.5%, specificity 92.72%, PPV 96.07%, and NPV 78.46%. Peak enhancement, wash-in area under the curve, wash-in rate, and the wash-in perfusion index were significantly different between the groups. No significant differences were found between rise time, mean transit time, and time to peak. For the ANN, sensitivity was 94.64%, specificity 94.44%, PPV 97.24%, and NPV 89.47%. LIMITATIONS Only PC and CP lesions were included. CONCLUSION Parameters obtained through TIC analysis can differentiate between PC and CP cases and can be used in an automated computer-aided diagnostic system with good diagnostic results. ( CLINICAL TRIAL REGISTRATION NUMBER NCT01315548.).