Stratifying preventive strategies by patient risk modifies the observed effectiveness against healthcare-associated infections: transfer matrices

This dataset consists of transfer matrices for patients stratified by their susceptibility to healthcare-acquired infections, distinguishing a group of patients visiting the hospital once in a considered period, for Lower Saxony-- a region of Germany.The matrices were based on anonymised healthcare records. These matrices are computed separately for each patient group (low- and high-risk groups and one-time patients). Patients are assigned to a risk group due to ICD-10 disease codes included in their hospitalization data, for the particular ICD-10 disease codes assigning patients to a high-risk group see [https://doi.org/10.1371/journal.pcbi.1008600, https://doi.org/10.1038/s41598-023-45248-3].The records used to estimate network transfer matrices were provided by a German health insurance company AOK Lower Saxony. These records spanned dates between January 1st, 2008 and December 31st, 2015. The records focused mainly on facilities situated in Lower Saxony and there were not enough records to accurately represent other federal states. Additionally, in our analysis, we excluded facilities that had no hospitalisations for at least 90 consecutive days. We also excluded records with missing diagnosis codes. After removing incomplete entries and entries corresponding to other regions, there were 4,223,023 entries of 1,419,712 distinct patients in 162 facilities. In the filtered data set we identified 181,914 high-risk patients (12.81% of all patients), 675,167 low-risk patients (47.57%) and 562,631 one-time patients (39.62%).Detailed analysis of the entire dataset and the subset for Lower Saxony only can be found in [https://arxiv.org/abs/1903.04701].To estimate network transfer matrices using patient's hospitalisation records,we determined each patient's transfers. As described in [https://arxiv.org/abs/1903.04701], deducing transfers was not always straightforward and required an algorithm to deal with non-standard cases. The detailed description of the algorithm developed for this purpose can be found in [https://doi.org/10.1371/journal.pcbi.1008442]. For low- and high-risk patients, we summed up the number of transfers between any two nodes and computed the average lengths of stay in each network node. Following, the step procedure outlined in [https://doi.org/10.1371/journal.pcbi.1008442] we constructed the transfer probability matrices, which were row-stochastic of size 2n*2n, with n denoting the number of hospitals in the network and in which the first n indices denote hospital nodes, while the latter n indices -- corresponding community nodes. An element (i,j) of the transfer probability matrix describes the probability of a patient transfer from node i to node j per day. For one-time patients we computed the average length of stay in hospitals, however, we were unable to compute the length of any community node stay, as they were defined as a period between two subsequent hospitalisations. Thus, we set the average length of stay in these nodes to 10 years, which agreed with the assumption that one-time patients visited a hospital only once during an eight-year-long period. Then, using the stochasticity condition, we followed the same procedure to create the transfer probability matrix. Due to the definition of the one-time patient group, this matrix was made of four diagonal n*n blocks.More details on the model corresponding to these matrices and interpretation of the community nodes may be found in [https://doi.org/10.1371/journal.pcbi.1008442, https://doi.org/10.1016/j.epidem.2020.100408, https://doi.org/10.1016/j.cmi.2022.08.001].