Mathematical and computational models play an increasingly important role in the assessment and control of epidemics. In addition to studies focusing on the general population, models are also used to study the spread of infectious diseases in specific settings, such as, e.g., hospitals. Nosocomial infections represent a major public health issue with high morbidity and mortality and high costs associated with prolonged treatments, which deserve urgent and efficacious prevention strategies. The sources and transmission paths of nosocomial infections are indeed often unknown and unrecognized. The application of epidemic models to hospitals is therefore crucial to provide valuable insights on the routes of infection propagation and to identify tailored measures for prevention and control of hospital acquired infections. Models need to be informed with the pattern of interactions among individuals along which the transmission of infection can occur, however only simple homogeneous assumptions have been considered so far. Our study provides for the first time direct measures of the number and duration of close contacts by different role and at the individual level in a hospital setting. By taking advantage of the RFID technology, our results account for important heterogeneities in the hospital population and in the interactions among patients, health care workers, and visitors, that enable an accurate parameterization of models for infectious disease spread on the close-contact route.
Our main finding is the very limited interaction that we observed between pairs of patients or between pairs of caregivers, and between health care workers and caregivers. This is empirically found both in the number of contact events, taking into account both distinct and repeated events, and in the duration of such contacts. This result has immediate practical implications for the development of prevention measures for respiratory infections within the hospital, which represent the most frequent nosocomial infections 
. Current guidelines 
identify the caregiver class as the priority group that control strategies should target 
, given that they may carry asymptomatic or mild community acquired respiratory illnesses and then spread the infection within the hospital to susceptible patients and staff. While caregivers may represent a source of the infection, our results show that their pattern of interaction is very stable and mostly spent in contact with the corresponding patient, thus not favoring the spread of a potential infection to a large number of individuals in the ward. This observation can inform models aimed at testing different prevention recommendations, exploring control resources focused mainly on the caregiver-patient interaction as opposed to resources focused on all possible interactions that a caregiver may have in the ward.
Another major insight concerns patients. Our results show that in addition to intense and continuative interactions with the caregivers, patients are contacted most frequently and with the longest duration by nurses, among all health care categories. Nurses were also found to have a pattern of frequent and long contacts among each other. Thus, in our setting, where a high proportion of patients had a diagnosis of respiratory infection, the intensive professional contact of nurses with patients and among themselves may result in a higher risk for airborne infections among nurses. These observations highlight the crucial importance of prioritizing nurses in local infection control interventions, and confirm the findings of a recent study where data on contacts were collected through a questionnaire 
. On the other hand, the pattern of contacts between physicians and patients showed a small number of contacts of short duration suggesting a less important role of this health care category in infection transmission. It is worth to mention that, although our study was conducted during the peak of influenza A/H1N1v activity in Italy, no case of influenza transmission was observed among the individuals included in the study. This observation is in line with a strict application of H1N1 containment guidelines in the hospital setting established by the Italian Ministry of Health 
We have studied the statistics associated with the collected data and we found a large degree of heterogeneity in the number of contacts and in the duration of contacts across classes. Large fluctuations up to 2 or 3 orders of magnitude are observed in the number of contacts established by an individual in a given class with any other individual, and similar results are observed in the duration of contacts. These fluctuations are however reduced when considering specific class-class interactions, highlighting the presence of well-defined interaction behaviors that are class-specific. Large variations are observed in the patient-caregiver interaction, providing additional empirical evidence for the importance of focusing control efforts on such interactions as they may lead to large variations in a potential outbreak.
Though observed in the hospital context, the large variability observed in the contact number and contact duration is consistent with empirical data collected in other settings. As an example, the same RFID technology was applied to investigate contact patterns in scientific conferences 
showing also in this case a strong heterogeneity of the contact durations between individuals – most contacts were very short in this setting, but contacts of very different durations were observed, including very long ones. This similarity points to the presence of common statistical signatures in the way people interact, that go beyond the constraints and behavioral patterns imposed by the specific context.
Our study allows to define risks of transmission between classes of individuals by calculating matrices of contacts of class-class type along the three quantities describing a contact event that we considered – the number of contacts, the number of distinct contacts, the overall duration of contacts. These contact matrices represent the input ingredients for the parameterization of mathematical and computational models of nosocomial infections, going beyond simple homogeneous assumptions and simple structuring of the population into two classes – patients and health care workers – as in previous studies 
. This work complements similar efforts that focus on the community level 
, though our data collection method ensures a higher objectivity of the measure of contact events, and a higher resolution both in time and in space. By taking full advantage of this resolution, RFID technology can be used to inform increasingly complex models that require a finer classification of contacts by personal characteristics 
. Differently from survey methodologies, our method affords data collection and behavior characterization at the individual level, and thus may be used to inform agent-based modeling approaches. This could provide additional insights and uncover unexpected behaviors induced by the fluctuations observed even within the class-class structuring of the population.
However, in order to obtain reliable and statistically significant results from numerical simulations, longer and more extensive deployments should be considered to better characterize expected behaviors and the associated fluctuations. Although this study was carried out over a period of one week, it was conducted in a period when strict rules for infection control were applied because of the threat of cross infection with the 2009 A/H1N1 pandemic influenza virus. We do not believe that the conduction of the study itself has changed the behavior of health care workers, but longer deployments would in this respect be needed to assess the stability of the results over time.
A strength of the present study is the high participation rate of individuals in our experiment. We did not encounter major opposition in the acceptance of the experimental procedure by health care workers, patients, or caregivers, and only a very limited fraction of them (3.6%) declined participation. Once enrolled in the study, most participants regularly wore the RFID devices according to the instructions, thus contributing to the collection of high quality data. These observations suggest that if appropriate protocols are provided, and privacy protection is ensured, measures of contacts through RFID devices are easily replicable, and it is conceivable to aim at fully covering entire wards or hospitals. In this respect, such an approach would constitute a major improvement in the collection of high quality data, if compared to available studies based on interviews where the participation rate was much lower and where logistics and resources limited the length and coverage of the survey 
Our method presents limitations as well. First of all, given that the RFID tags exchange radio-frequency signals, the collected data can only provide information on the proximity of two badges (and therefore of the persons carrying the badges), but no information on the possible occurrence of a physical contact between the two persons is available. Our measures can thus be used for properly estimating the transmission parameters of respiratory infections but they are less informative for infections transmitted by direct contact. Note however that physical contact can only occur between persons who are already in spatial proximity. Therefore, it would be very interesting to study the fraction of close encounters that result in a physical contact, as this may help to identify useful parameters for modeling infections transmitted through physical contact. In the future, the use of sensors that can directly resolve physical contact may be explored. Another possible limitation is the fact that settings with no wireless connections involve additional complexity in the implementation of the measurements and in the definition of procedures, as limited by the available communication infrastructure. However, upcoming new technology that will allow operating the RFID sensing layer in a fully distributed fashion with on-board storage on the devices, will have minimal requirement on the host infrastructure and is expected to provide increasingly larger opportunities for deployment of wearable sensing systems for the measurement of contacts patterns. This new technology would be also very important to run similar data collection campaigns that cover full hospitals, to assess whether logistic and behavioral characteristics and procedures that are ward-specific are reflected in the observed contact patterns. Finally, a comparative study between different hospitals may provide valuable insights into how structural organization and procedural management may impact the contact patterns among individuals, and therefore the potential epidemic spreading within the hospital. This would also allow us to assess the specificity of the results obtained in the present study. On the other hand, human behavior has been shown in many studies to exhibit important regularities, and a certain number of characteristics of our results can be expected to hold across different wards or hospitals: for example the broad distributions of , or the overall structure of the interaction networks, with the strong specificity of patient-caregiver contacts contrasted to the HCW interactions.
Our study represents, to our knowledge, the first example of unsupervised data collection of face-to-face contacts in a hospital setting by means of wearable radio frequency devices. The obtained results provide significant advances in our knowledge of the mixing patterns taking place in a hospital ward, and allow for a fine structuring of the population into classes of individuals based on their role, along with the evaluation of the corresponding contact matrices. The resulting analysis may help to identify specific interactions at increased risk of transmission, and to explore a variety of possible interventions by means of numerical simulations obtained with modeling approaches informed by measured contact matrices. Outbreak investigations conducted concurrently with proximity sensing by wearable devices may further augment the knowledge we have on the routes of transmission and thus help in reducing the burden of nosocomial infection.