The automated malnutrition assessment

Objective: We propose an automated nutritional assessment algorithm that provides a method for malnutrition risk prediction with high accuracy and reliability.

Methods: The database used for this study was a file of 432 patients, where each patient was described by 4 laboratory parameters and 11 clinical parameters. A malnutrition risk assessment of low (1), moderate (2), or high (3) was assigned by a dietitian for each patient. An algorithm for data organization and classification using characteristic metrics for each patient was developed. For each patient, the algorithm characterized the patients' unique profile and built a characteristic metric to identify similar patients who were mapped into a classification. For each patient, the algorithm characterized the patients' classification.

Results: The algorithm assigned a malnutrition risk level for different training sizes that were taken from the data. Our method resulted in average errors (distance between the automated score and the real score) of 0.386, 0.3507, 0.3454, 0.34, and 0.2907 for the 10%, 30%, 50%, 70%, and 90% training sizes, respectively. Our method outperformed the compared method even when our method used a smaller training set than the compared method. In addition, we showed that the laboratory parameters themselves were sufficient for the automated risk prediction and organized the patients into clusters that corresponded to low-, low-moderate-, moderate-, moderate-high-, and high-risk areas. The organization and visualization methods provided a tool for the exploration and navigation of the data points.

Conclusion: The problem of rapidly identifying risk and severity of malnutrition is crucial for minimizing medical and surgical complications. These are not easily performed or adequately expedited. We characterized for each patient a unique profile and mapped similar patients into a classification. We also found that the laboratory parameters were sufficient for the automated risk prediction.