Dynamic Contact Network Analysis in Hospital Wards

We analyse a huge and very precise trace of contact data collected during 6 months on the entire population of a rehabilitation hospital. We investigate the graph structure of the average daily contact network. Our main results are to unveil striking properties of this structure in the considered hospital, and to present a methodology that can be used for analysing any dynamic complex network where nodes are classified into groups. The MOSAR project aims at examining the factors determining the dynamics of AMRB (AntiMicrobial Resistant Bacteria) spread within healthcare facilities. To further reduce transmission, in addition to classical prevention measures (such as admission controls, isolation of carriers and hand hygiene), changing contacts within the hospital is considered as the next step [1]. Indeed, contacts strongly influence how transmission occurs [2]. Yet, contacts are difficult to measure efficiently in practice, and they may even be harder to change. Recently, however, advances in communication technologies have made it possible to record personto-person interactions with unprecedented detail, allowing an in depth view of the structure of contacts in real-life settings [3]. If such contacts actually support transmission, it may open the way to further improvement in hospital hygiene. In this article, we analyse the contact trace recorded on the entire population of a rehabilitation hospital during 6 months between June and November 2009, within the MOSAR project. We focus on a period of 8 weeks of the measurement, from July 6th to September 2nd involving 492 individuals, 253 patients and 239 staffs. We describe the methodology we used to uncover the key characteristics of this dynamic contact network and the main results we obtained: we point out big differences in the contact profiles of services (Sec. 1), as well as in contact patterns of patients and staffs (Sec. 2), and we reveal the structure of interconnections between the mainly introverted services of the hospital (Sec. 3). Related works There have been some works using sensor devices in order to unfold contact patterns among individuals in environments involving patients or children, which present critical risks for spreading of diseases. The measurement analysed in [4] was made on an entire primary school during 3 days. Two similar experiments, described in [5,6], were both conducted during one week in some paediatric ward. Compared to those works, our analyses present two important advantages. Firstly, the measurement we use was made on a much longer period of time (6 months), which allows to assess the generality of the conclusions we can derive on shorter period of times (like one day or one week). Secondly, our measurement is not limited to a specific part of the hospital, it involves all patients and all staffs of all services of the hospital, which is a key point to have an accurate view of the actual possibility of spreading into a given service. Indeed, these possibilities highly depend on the contacts occurring outside the service under study. Preliminaries The contact data was recorded using sensor devices carried by the participants and that send signals every 30s. Those signals include the ID of their source device which is recorded together with a time stamp by devices that are close enough from this source (typically 1 to 2 meters). The sending time of the different sensors are not synchronised but their internal clocks are. Afterwards, time is sliced in slots of 30s and we keep, for each slot, the list of pairs A,B of sensors such that at least one (possibly both) recorded the signal of the other. Each of these pairs is unordered (we do not keep track of which node receives the signal and which one sends it) and appears at most once in a given time slot. Finally, in all this article, we manipulate intervals of contacts instead of punctual contacts, i.e. a contact is a quadruplet (A,B, ts, te) where A and B are two nodes of the network and ts and te are respectively the time slots where starts and ends the interval of contact between A and B, the length of the contact being te − ts. Throughout the article, we analyse sets of contacts over a specified time period (typically one day) using three parameters: number of contacts, cumulated length of contacts and number of adjacency pairs. (A,B) is an adjacency pair on a given time period iff there is at least one contact between A and B during this period. A contact (A,B, ts, te) between A and B gives rise to two semi-contacts: one attached to A, denoted (A, ts, te) B , and one attached to B, denoted (B, ts, te) B . And similarly, every adjacency pair gives rise to two adjacency semi-pairs. In the rest of the article, for sake of simplicity of vocabulary, we use the term contact (resp. pair) instead of semi-contact (resp. semi-pair), but all statistics are actually made using semi-contacts (resp. semi-pairs). The reason is that it gives a straightforward meaning to mean statistics per individual. In the rest of the article, for sake of comparison, we make extensive use of a uniformised version of the network of the hospital, which we call the full-uniform network and which is defined as follows. The full-uniform network is a complete weighted graph where each pair of nodes receives 1) a weight equal to the density of the real network (i.e.the mean number of adjacency pairs per pair of nodes in the real network), 2) a number of contacts equal to the mean number of contacts per pair of nodes, and 3) a cumulated length of contacts equal to the mean cumulated length per pair of nodes. 0 5 10 15 20 25 30 35 S1 S2 S3 S4 S5 S6 S7 S8 S9 Nu m be r o f i nd iv id ua ls Services Staffs Patients Fig. 1: Number of individuals per day for each service, distinguishing between patients (light red) and staffs (deep blue). General organisation of the hospital Over the period of study, the mean number of people present in the hospital in one day is about 103 patients and 64 staffs. The patients and staffs are divided into 9 services (see repartition on Fig. 1), only the first five of which (S1 to S5) contain both patients and staffs, the other four (S6 to S9) containing only staffs. Each of the services S1 to S5, containing both patients and staffs, occupies one floor in one of the two wings of the building: S1, S2 and S3 occupy respectively the 1st, 2nd and 3rd floor of the 1st wing, while services S4 and S5 occupy the 2nd and 3rd floor of the 2nd wing. Services S7 to S9 contain rehabilitation staffs and S6 is the night service, regrouping people replacing staffs from services S1 to S5 during nights. S7 and S8 are located in two distinct places between the two wings of the buildings, but S6 and S9 do not have a unique location in the hospital. It must be clear that the division of the hospital into services is not meaningful only from an administrative point of view but has also a strong impact on the structure of the network: in average in one day, 66% of the adjacency pairs of the hospital occur inside services, and 92% of the cumulated length of contacts, while these values are only 25% in the full-uniform network. 0 100 200 300 400 500 600 S1 S2 S3 S4 S5 S6 S7 S8 S9 Av er ag e de gr ee p er s er vi ce p er d ay Services