TIAMAT- towards an interdisciplinary automated malnutrition screening tool

Abstract

Background

Malnutrition remains a major problem in the world's hospitals, and although nutrition screening (such as by Nutrition Risk Screening NRS-2002) improves detection and management, it has proved difficult to ensure that this is routinely performed. There is a case for automated approaches that are independent of staff time. As common laboratory tests such as serum albumin are not considered markers of malnutrition, they are not included in screening tools, but it is possible that combinations of tests would predict malnutrition, and probable that high risk patients would be identified.

Methods

We studied 300 unselected consenting internal medicine patients recently admitted to a university hospital. We determined NRS-2002, Malnutrition Universal Screening Tool (MUST) and Subjective Global Assessment (SGA). No additional interventions were performed. The laboratory database was searched for investigations already performed. The dataset was split randomly into 200 for training and 100 for validation. The primary endpoint was an algorithm to predict SGA (Towards Interdisciplinary Automated MAlnutrition screening Tool - TIAMAT). Blood results for which there were data from at least 60 % of patients were selected; dimensionality was checked with factor analysis, and a multivariate logistic model using stepwise regression was formed. The predictions of SGA from NRS-2002 and MUST were compared with those from the newly created score.

Results

Simple haematological and biochemical tests had been performed in all patients. To predict SGA (none vs moderately/severely malnourished) the training set yielded a score with an optimal sensitivity of 71 % (95 % CI 58–81 %), specificity 81 % (70–89), positive predictive value (PPV) 79 % (66–88) negative predictive value (NPV) 73 % (62–83) and AUC 0.81 (0.74–0.88). In the validation cohort - in which missing data were imputed with the relevant median – TIAMAT had sensitivity of 60 % (44–75) specificity of 75 % (62–86), PPV 65 % (48–79), NPV 72 % (59–83) and AUC 0.77 (0.69–0.82) compared with MUST ≥2 for which sensitivity was 63 % (47–77), specificity 91 % (81–97), PPV 84 % (67–95), NPV 76 % (65–86) and AUC 0.80 (0.72–0.89), and NRS-2002 for which sensitivity was 93 % (81–99), specificity 91 % (81–97), PPV 89 % (76–96), NPV 95 % (85–99) and AUC 0.92 (0.86–0.97).

Conclusion

Composites of standard laboratory data form a potential alternative to current screening methods, the results of which could be displayed to all data viewers without any human intervention. This could facilitate and thus improve the efficacy of screening by extending it to the entire hospital population.