Using machine learning to predict nosocomial infections and medical accidents in a NICU

Abstract

Background

Adult studies have shown that nursing overtime and unit overcrowding are associated with increased adverse patient events but there exists little evidence for the Neonatal Intensive Care Unit (NICU). We investigate the main determinants of nosocomial infections and medical accidents in a NICU using state-of-the-art machine learning techniques. Our analysis focuses on a retrospective study on the 7,438 neonates admitted in the CHU de Québec NICU (capacity of 51 beds) from 10 April 2008 to 28 March 2013. Daily administrative data on nursing overtime hours, total regular hours, number of admissions, patient characteristics, as well as information on nosocomial infections and on the timing and type of medical errors were retrieved from various hospital-level datasets.

Methods

We use a generalized mixed effects regression tree model with random effects (GMERT-RI) to elaborate predictions trees for the two outcomes. Neonates' characteristics and daily exposure to numerous covariates are used in the model. GMERT-RI is suitable for binary outcomes and is a recent extension of the standard tree-based method. The model allows to determine the most important predictors.

Results

Diagnosis-related group level, regular hours of work, overtime, admission rates, birth weight and occupation rates are the main predictors for both outcomes. On the other hand, gestational age, C-Section, multiple births, medical/surgical and number of admissions are poor predictors.

Conclusion

The GMERT-RI algorithm is a powerful tool. It is well suited to unearth potential correlations in the context of unbalanced panel data and discrete health outcomes, two common features of clinical data. In the particular setting of a NICU, we find that institutional features (overtime hours, occupancy rates, etc.) are just as important drivers as neonate-specific medical conditions in predicting medical accidents and health care associated infections. From an operational point of view, prediction trees can complement traditional management tools in preventing undesirable health outcomes in the NICU.

Appendix: the GMERT-RI algorithm

The GMERT-RI algorithm of [13] is defined as follows. Recall that for the generalized mixed model (GLMM),

• $y_{it} \mid u_i$ belongs to the exponential family of distribution.

• $\mu _{it}=E \left[ y_{it} \mid u_i \right]$ and $g \left( \mu _{it} \right) = \eta _{it} = X_{it} \beta + Z_{it} u_{i}$ for some known link function g. In our case, we use the logit link. $\mu _{it} = \frac{e^{ \eta _{it} }}{1 + e^{ \eta _{it} }}$ and $g \left( \mu _{it} \right) = \log \left( \frac{\mu _{it} }{1 - \mu _{it} } \right) = \eta _{it}$. So, $\mu _{i}=E \left[ y_{i} \mid u_i \right]$ and $g \left( \mu _{i} \right) = \eta _{i} = X_{i} \beta + Z_{i} u_{i}$ with $u_{i} \sim N \left( 0, \Sigma \right)$.

• $Cov \left[ y_{i} \mid u_i \right] =\sigma ^2 v \left( \mu _{i} \right)$ where $\sigma ^2$ is a dispersion parameter and $v \left( \mu _{i} \right) = diag \left[ v_{i1}, ..., v_{iT_i} \right] = diag \left[ v \left( \mu _{i1} \right) , ..., v \left( \mu _{iT_i} \right) \right]$ where $v \left( . \right)$ is a known variance function.

The generalized mixed effects regression tree (GMERT-RI) model, proposed by [13], can be written as $\eta _{i} = f\left( X_{i} \right) + Z_{i} u_{i}$ with $u_{i} \sim N \left( 0, \Sigma \right)$ where the linear fixed part $X_{i} \beta$ is replaced by the function $f\left( X_{i} \right)$ that will be estimated with a standard regression tree model. A first-order Taylor-series expansion yields the linearized response variable, $\widetilde{y}_i = g \left( \mu _{i} \right) + \left( y_i - \mu _i \right) g^{\prime } \left( \mu _{i} \right)$ and the mixed fixed effect regression tree (MERT) pseudo-model is defined as follows: $\widetilde{y}_i = f\left( X_{i} \right) + Z_{i} u_{i} + e_i$. The GMERT-RI algorithm is basically the penalized quasi-likelihood (PQL) algorithm used to fit GLMMs where the weighted linear mixed effects (LME) pseudo-model is replaced by a weighted MERT pseudo-model. Therefore, the fixed part $f\left( X_{i} \right)$ is estimated with a standard regression tree model. The GMERT-RI algorithm of [13] is the following:

Algorithm 1

GMERT-RI ALGORITHM (see Hajjem et al. [13], p. 115)

As shown by [13], the GMERT-RI model can be used to predict the response for two categories of new observations: those who belong to a cluster included in the sample used to fit the model and those excluded from the sample. To predict the response for a new observation from the first category, one uses both its corresponding fixed component prediction $\widehat{f}\left( X_i \right)$ and the predicted random part $Z_i \widehat{u}_{i}$ corresponding to its cluster. This is a cluster-specific estimate. For the latter category, one can only use its corresponding fixed component prediction (i.e., the random part is set to 0).