A Large-Scale Exploration of Factors Affecting Hand Hygiene Compliance Using Linear Predictive Models

A Lar ge-Scale Exploration of F actors Af fecting Hand Hygiene Compliance Using Linear Predicti v e Models Michael T . Lash ∗ MS, Jason Slater † BS, Philip M. Polgreen ‡ MD MPH, and Alberto Maria Segre ∗ PhD ∗ Department of Computer Science, The Univ ersity of Iow a, Iowa City , IA 52242, Email: { michael-lash, alberto-segre } @uio wa.edu † Gojo Industries, Inc., Akron, OH 44309-0991, Email: slaterj@gojo.com ‡ Department of Internal Medicine, The Univ ersity of Iow a, Iowa City , IA 52242, Email: philip-polgreen@uio wa.edu Abstract —This large-scale study , consisting of 24.5 million hand hygiene opportunities spanning 19 distinct facilities in 10 different states, uses linear predicti ve models to expose factors that may affect hand hygiene compliance. W e examine the use of featur es such as temperature, relativ e humidity , influenza severity , day/night shift, federal holidays and the presence of new r esidents in pr edicting daily hand hygiene compliance. The results suggest that colder temperatures and federal holidays hav e an adverse effect on hand hygiene compliance rates, and that individual cultur es and attitudes regarding hand hygiene exist among facilities. Index T erms —Public healthcare, Hand hygiene, Supervised learning, Linear regression, Event detection I . I N T RO D U C T I O N Healthcare associated infections represent a major cause of morbidity and mortality in the United States and other countries [1]. Although many can be treated, these infections add greatly to healthcare costs [2]. Furthermore, the emer- gence of multidrug resistant bacteria hav e greatly complicated treatment of healthcare associated infections [3], making the prev ention of these infections ev en more important. One of the most effecti ve interventions for prev enting healthcare associ- ated infections is hand hygiene [4]. Y et, despite international programs aimed at increasing hand hygiene [4], [5], [6], rates remain low , less than 50% in most cases [4], [6], [7]. Because of the importance of hand hygiene in prev enting healthcare associated infections, infection control programs are encouraged to monitor rates to encourage process im- prov ement [6], [8], [9]. In most cases, hand hygiene mon- itoring is done exclusiv ely by human observers, which are still considered the gold standard for monitoring [7]. Y et, human observations are subject to a number of limitations. For example, human observers incur high costs and there are difficulties in standardizing the elicited observ ations. Also, the timing and location of observers can greatly af fect the di versity and the quantity of observations [10], [11]. Furthermore, the distance of observers to healthcare workers under observation and the relativ e busyness of clinical units can adversely affect the accurac y of human observers [11]. The presence of human observers may artificially increase hand hygiene rates tem- porarily as the presence of other healthcare w orkers can induce peer effects to increase rates [12], [13]. Finally , the number of human observ ations possible is quite small in comparison to the number of opportunities [7], [12]. Se veral automated approaches to monitoring hav e been proposed [8], [14], [15], [16]. Many of these measure hand hygiene upon entering and leaving a patient’ s room. The subsequent activ ation of a nearby hand hygiene dispenser is recorded as a hand hygiene opportunity fulfilled whereas, if no such acti vation is observ ed, the opportunity is not satisfied. Such approaches, while not capturing all fi ve moments of hand hygiene, do provide an easy and con venient measure of hand hygiene compliance. W ith automated approaches becoming more common, a more comprehensiv e picture of hand hygiene adherence should emerge, providing new insights into why healthcare workers abstain from practicing hand hygiene. I I . D A TA A N D M E T H O D S A. Hand Hygiene Event Data Our hand hygiene event data is a proprietary dataset pro- vided by Gojo Industries. The data were obtained from a number of installations consisting of door counter sensors , which increment a counter anytime an individual goes in or out of a room, and hand hygiene sensors , which increment a counter when soap or alcohol rub are dispensed. Additional supporting technology was also installed to collect and record timestamped sensor -reported counts. In this paper , we will use the term dispenser event to designate triggering and use of an instrumented hand hygiene dispenser and door event to designate the triggering of a counter sensor located on one of the instrumented doors. A total of 19 facilities in 10 states were outfitted with sensors; because of pri vac y concerns, we provide only the state and CDC Division for each. The facilities comprise a wide range of geographies, spanning both coasts, the midwest, and the south. A total of 1851 door sensors and 639 dispenser sen- sors reported a total of 24,525,806 door ev ents and 6,140,067 dispenser e vents, beginning on October 21, 2013 and ending on July 7, 2014. Each facility contributed an average of 172.3 r eporting days , making this study the largest inv estigation of hand hygiene compliance to date (i.e., larger than the 13.1 million opportunities reported in [17]). Assuming each door ev ent corresponds to a hand hygiene opportunity , we compute an estimated intra-facility compliance rate of 25.03%, in line with if not just below the reported low-end rate found in [18]. The original data, consisting of timestamped counts reported from individual sensors over short intervals, were re-factored to support our analysis. First, data from each sensor were binned by timestamp, t , into 12 hour interval s, corresponding to traditional day and night shifts, as indicated by an additional variable, nig htS hif t , defined as follows: nig htS hif t = ( 1 t  [7 pm , 6 : 59 am ] 0 t  [7 am , 6 : 59 pm ] Second, door and dispenser counts were aggregated based on day and night shift so as to produce a series of records. For each such record we compute hand hygiene compliance , or just compliance , by di viding the number of reported dispensed ev ents by the number of door ev ents: compliance = # dispenser # door Such a definition of compliance assumes that each door ev ent corresponds to a single hand-hygiene opportunity and each dispenser ev ent corresponds to a single hand-hygiene event whereas, in reality , a health care work er might well be expected to perform hand hygiene more than once per entry , resulting in rates that exceed one, if only slightly . This estimator also ignores the placement of doors with respect to dispensers: multiple dispensers may well be associated with a single doorway , and some dispensers may be in rooms having multiple doors. Adding new dispensers will raise apparent compliance rates, while adding new door sensors will appear to reduce compliance. Even so, when applied consistently and if system layouts are fixed, this estimator is a reasonable approximation of true hand hygiene compliance, and supports sound comparisons within a facility (but not across facilities). Because malfunctioning sensors or dead batteries can pro- duce outliers (i.e., very low or very high values), records with fewer than 10 door or dispenser events reported per day (possibly indicating an installation undergoing maintenance), zero compliance, or compliance values greater than 1 were remov ed prior to analysis (at the cost of possibly excluding some legal records). The remaining data consists of 5308 shifts from the original 5647 records, having 21,273,980 hand hygiene opportunities and 5,296,749 hand hygiene ev ents (see T able I). B. F eature Definitions In this subsection we define the features (factors) that will be examined, and how each is derived. Facility State CDC Div T ot Disp T ot Door Days Rep 91 OH ENC 234292 518772 252 101 OH ENC 350901 2021665 260 105 TX WSC 238899 1940024 260 119 MN WNC 123877 242939 156 123 TX WSC 325618 1112198 243 127 NM Mnt 1306855 4546171 260 135 OH ENC 125731 264331 258 144 CA Pac 398961 1744642 260 145 CA Pac 567096 2073566 260 147 CA Pac 500979 2462900 260 149 CA Pac 590708 2306392 260 153 CT New E 169564 603482 208 155 NY M-At 171275 619507 117 156 NC S-At 4381 38200 15 157 OH ENC 39455 313396 101 163 OH ENC 344 10233 5 168 P A M-At 30421 86909 20 170 IL ENC 112604 353631 47 173 OH ENC 4788 15122 32 T otal 10 8 5296749 21273980 3274 T ABLE I: Descriptiv e statistics for all reporting facilities in terms of state, CDC division, hand hygiene ev ents, people ev ents, and reporting days. 1) Local W eather Data: Because health care workers fre- quently cite skin dryness and irritation as a factor in decreased compliance (particularly in cold weather months where en- vironmental humidity is reduced), we associate daily air temperature and relative humidity to each timestamped record based on each f acility’ s reported zip code. Spatially assimilated weather v alues ( σ = 0 . 995 ) for the entire globe were obtained from the National Oceanic and Atmospheric Administration (NO AA) [19]. Given in terms of grid elements (a tessilation of bounding boxes covering 2 . 5 ◦ latitude by 2 . 5 ◦ longitude), the world is thus defined as a 144 by 73 grid having 10512 distinct grid elements. W eather data are av ailable at a fine lev el Fig. 1: Assigning ( red box) NOAA weather data, reported in terms of a geographic grid, to health care facilities ( red dots), where the blue color gradient might represent temperature. of temporal granularity (on the order of 4 times daily for each grid unit) for the entire period of interest. The geographical assignment of weather data was obtained by first mapping each facility’ s numerical zipcode to the zipcode’ s centroid (2010 US Census data), giv en by (latitude,longitude), which was subsequently mapped to the matching NO AA grid element. An example of this assignment can be observed in Figure 1. W e associate weather information from the 6am reporting hour with records corresponding to traditional day shifts (7am- 6:59pm) and use the 6pm reporting hour for traditional night shifts (7pm-6:59am). 2) Influenza Severity: W e conjecture that the local severity of common seasonal diseases, such as influenza, may also affect hand hygiene compliance rates. W e define influenza sev erity as the number of influenza-related deaths relativ e to all deaths ov er a specified time interval. Influenza sev erity data were obtained from the CDC’ s Morbidity and Mortality W eekly Report (MMWR), which also reports data at weekly temporal granularity . Rather than reporting data by CDC region, howe ver , data are provided by r eporting city (one of 122 participating cities, mostly large metropolitan areas). W e map each facility in our dataset to the closest reporting city in order to associate the appropriate sev erity value to each record. In other words r epC ity = argmin { dist ( facility , city i ) : i = 1 , . . . , 122 } where dist ( fac , city ) , k ( fac lat , fac lon ) , ( city lat , city lon ) k 2 , the Euclidean distance between two entities given in terms of (lat, lon) coordinates. Eight of 19 facilities were located in a reporting city (i.e., dist = 0 ). The remaining 11 facilities were mapped to a reporting city that was, on av erage, 66.2 miles away (only 3 of 19 facilities were mapped to a reporting city further than this a verage, with the largest distance being 142 miles). 3) T emporal F actors: W e also conjecture that external factors associated with specific holidays or ev ents may affect hand hygiene compliance rates. Holidays may change staffing rates or af fect healthcare w orker behaviors in various ways. The number of visitors (af fecting door counter rates) may also be greater than during regular weekdays. Holidays such as the 4th of July are often associated with alcohol-related accidents, and may increase health care facility workloads. Such factors may also be observable during weekends. W e define a new v ariable hol iday that reflects whether a giv en shift occurs on one of the 10 federal holidays (New Y ear’ s Eve, Martin Luther King Day , President’ s Day , Memo- rial Day , the 4th of July , Labor Day , Columbus Day , V eteran’ s Day , Thanksgiving or Christmas): holiday = ( 0 t / ∈ { hol idays } 1 t ∈ { holiday s } Similarly , in order to ascertain the impact of weekends on compliance, we define a new variable w eekday as follows: w eekday = ( 0 t ∈ { S at, S un } 1 t ∈ { M on, T ues, W eds, T hurs, F ri } A related concept is the presence of new resident physicians, who traditionally start work the first of July . W e define a new variable that corresponds with this time period in order to see if the data rev eal the presence of a July effect: J uly E f f ect = ( 0 t / ∈ J ul y 1 − 7 1 t ∈ J uly 1 − 7 C. Exploring F actors Affecting Hand Hygiene 1) M 5 Ridge Regr ession for F eature Examination: With cov ariates defined and associated with the collected sensor data, we wish to build a linear hypothesis h that (a) accurately estimates hand hygiene and (b) reports the direction and degree of effect of our defined features. In accomplishing (b) we bear in mind two things: (1) There may be multi-collinearity among features, which may adversely affect the output. (2) That (a) and (b) may be at odds with one another; i.e., obtaining good predictions may entail discarding some prediction-inhibiting features for which we would like to obtain ef fect estimates (in practice, we find that this is not actually the case). Therefore, we propose an M 5 Ridge Re gr ession for F eatur e Examination method designed to accomplish (a) and (b) , while bearing (1) and (2) in mind. This method is giv en by h ∗ = argmin h ∈H l k Λ( X ) h − y k 2 2 + λ k h k 2 2 s.t. ρ ( h j ) ≤ . 05 ∀ j (1) where X ∈ R n × p is a design matrix, h is the hypothesis, y is the target vector consisting of compliance rates in which a particular y i ∈ [0 , 1] , λ is a regularization term, k·k 2 is the ` 2 -norm, and ρ ( · ) is a function that reports the p- value of a hypothesis term (this constraint is ensured via sequential backwards elimination [20]). The function Λ( X ) can be defined as Λ( X ) , argmin { t ∈ T H l } (2) where t is hypothesis selected from a tree of hypotheses constructed using the M 5 method [21]. Ef fectiv ely , (2) only reduces the p dimension, acting as a feature selection method, and having no bearing on the n dimension. There are a few benefits of the above method worth pointing out. First, the hypothesis class H l is linear and common to both (1) and (2). T wo-stage optimization approaches, where the first objectiv e is optimized, taking into account the hypothesis class, before the hypothesis itself is optimized for predictiv e accuracy (or some other such measure), ha ve been shown to work well [22]. Secondly , such a method is specifically geared tow ard producing a hypothesis that makes use of features that ha ve an immediate bearing upon the problem, while eliminating interpretability obscuring ef fects, such as multi- collinearity . Moreo ver , these desirables are obtained while attempting to produce the most accurate hypothesis: an h that elicits feature indicativeness, produces accurate results, and controls for confounding effects is the goal of this two-step optimization procedure. Ultimately , we conduct our analysis by observing the sign and magnitude of the values in the hypothesis vector in order to determine the factors that influence hand hygiene compli- ance, and whether such factors affect compliance in a positiv e or negati ve manner . W e also observe correlation and RMSE values to determine how well our predicti ve model works, and whether the corresponding results can be trusted. All results and are obtained via k -fold cross-validation ( k = 10 ). 2) Supporting Methodology: W e also use two estab- lished/standard techniques – RReliefF feature ranking and marginal effects modeling – that will serve as a point of comparison between our method, and also help inform the discussion of the obtained results 1 . Featur e ranking: First, we propose the use of the RReliefF algorithm [25], a modification of the original Relief algorithm of Kira and Rendell [26]. RReliefF finds a feature j ’ s weight by randomly selecting a seed instance x i from design matrix X and then using that instance’ s k nearest neighbors to update the attribute. This description consists of three terms: the probability of observing a different rate of hand hygiene compliance than that of the current v alue giv en that of the nearest neighbors, giv en by A = p ( rate 6 = rate x i,j | k NN ( x i,j )) , (3) the probability of observing the current attribute value gi ven the nearest neighbors, giv en by B = p ( x i,j | k NN ( x i,j )) , (4) and the probability of observing a different hand hygiene rate than the current value giv en a dif ferent feature v alue v and the nearest neighbors, giv en by C = p ( rate 6 = rate x i,j | k NN ( x i,j ) ∧ j = v ) . (5) Attribute distance weighting is used in order to place greater emphasis on instances that are closer to the seed instance when updating each term; final weights are obtained by applying Bayes’ rule to the three terms maintained for each attribute, which can be expressed C ∗ B A − (1 − C ) ∗ B 1 − A . (6) By using this method we could then rank attributes in terms of their importance. W e again report rankings using k -fold ( k = 10 ) cross validation. Marginal Effects Modeling: T o provide additional insight into the features that are relev ant to hand hygiene we analyzed 1 Note that both the LASSO [23] and Elastic Net [24] would have also made appropriate supporting methods. their marginal effects [27]. Marginal ef fects, also referred to as instantaneous rates of change , are computed by first training a hypothesis h , then, using the testing data, the effect of each cov ariate can be estimated by holding all others constant and observing the predictions. Such a method can be expressed by ˆ rate i,j = h > [ x i,j , ¯ x 6 = j ] (7) where, with a slight abuse of notation, x i,j , the value of instance i ’ s j th feature, is added to the vector ¯ x 6 = j , which consists of the av erage of each non- j feature, at the appropriate location (namely , the j th position). Here, the notation 6 = j is used to reinforce the fact that the vector of averages ¯ x has it’ s j th element replaced by x i,j . Other non- j entries are given by ¯ x k = µ ( X k ) , for an arbitrary index position k . I I I . R E S U L T S A. Pr edictive P ower: M 5 Ridge Re gression W e learned a hypothesis using all av ailable features, in- cluding a nominalized facility identifier . Our predictiv e results can be observed in T able II. W e note that the RMSE is not large and the correlation is moderate, implying relativ ely good predictiv e performance. Measure V alue Correlation 0.3441 RMSE 0.1702 T ABLE II: Correlation coefficient and RMSE of cross- validated model predictions. B. Examining Hypothesis h ∗ W e ne xt examine the terms of the learned hypothesis h ∗ (see T able III). The model includes all 19 facilities, 12 of which had positi ve values, indicating relatively higher rates of compliance. The size remaining facility’ s h ∗ terms had relativ ely small negati ve values, indicating lower rates of com- pliance. Among other features, holidays are associated with lower compliance rates, while influenza se verity has higher compliance. W eekdays are associated with higher compliance rates, as are higher temperatures and humidity . Interestingly , the M 5 Ridge Regression model appears to hav e eliminated Featur e h j Facility − = { 1 , 105 , 147 , 156 , 157 , 170 } h j ∈ Fac − ∈ [ − 0 . 103 , − 0 . 016] Facility + = { 91 , 119 , 123 , 127 , 135 , 144 , h j ∈ Fac + ∈ 145 , 149 , 153 , 155 , 168 , 173 } [0 . 008 , 0 . 261] Air T emp 0 . 022 Rel. Humid 0 . 0079 week day 0 . 0069 nig htShif t − 0 . 0218 holiday = { Indep Day , Pres. Day , h j ∈ Hol V et Day , New Y ear’s , Christmas } [ − 0 . 017 , − . 006] Flu Severity 0 . 014 J uly E f f ect − 0 . 0106 T ABLE III: Feature specific h j terms, where red highlights features with a negati ve association and blue highlights those with a positiv e association. some holidays (Martin Luther King day , Memorial day , Labor day , Columb us day , and Thanksgi ving), as well as Facility 163 (the facility with the lowest amount of hand-hygiene data). This means that these features do not contribute to hand- hygiene compliance rates in any meaningful way . C. RReliefF By using RReliefF we could rank features in terms of their importance in order to support and supplement the result obtained using M 5 Ridge Regression. The results are reported in T able IV, where rankings shown are averages for 10-fold cross-validation. Note that here f acility was represented as a single discretely-valued feature in order to determine the importance of facility as a whole (instead of treating each facility as its own feature), as was hol iday . Attribute A vg V al A vg Rank Facility 0 . 029( ± . 001) 1 Flu Sev 0 . 007 2 Air T emp 0 . 005 3 . 3( ± 0 . 46) week day 0 . 002 5 Rel. Humid. . 001 6 . 3( ± 0 . 64) J uly E f f ect ≈ 0 . 0 7 . 2( ± 0 . 4) holiday ≈ 0 . 0 7 . 8( ± 1 . 08) nig htShif t ≈ 0 . 0 8 . 7( ± 0 . 46) T ABLE IV: RReliefF attribute weights. D. Mar ginal Effects The results obtained from modeling the marginal effects can be observed in Figure 2. Figures 2a and 2b sho w the marginal effects of two ran- domly selected facilities; one identified as being associated with lower rates of compliance and one identified as having higher rates of compliance (from T able III). Note that, because these are binary features (taking on v alues of either zero or one), the kernel density of the underlying data is not readily visible (unlike the other figures, which sho w results for non- binary features). As we can see the mar ginal ef fects support the result obtained using both M 5 Ridge Regression and RReliefF , and also seem to suggest an even greater association between facilities and rates of compliance than was originally apparent (at least for these two facilities). Figure 2c shows the mar ginal effects of flu Se verity . The Flu Sev erity result shows a slightly positiv e relationship between the se verity of flu, measured in terms of mortality , and hand- hygiene compliance rates. This is further supported by the result obtained from M 5 Ridge Regression and the RReliefF ranking. Figures 2d and 2e show the marginal effects of humidity and temperature. The result obtained for both is consistent with that from M 5 Ridge Regression. The lesser effect of humidity and greater effect of temperature are also reflected in the RReliefF ranking. T o further explore the relationship between hand-hygiene and weather effects, we conducted a simple statistical analysis. For each facility , we selected the temperature and humidity values corresponding to the bottom 10% and top 10% of hand- hygiene compliance rates. W e then performed a paired t-test on each set of samples; temperature and humidity v alues were scaled to [0 , 1] . The results of this analysis are reported in T able V. Facility State T emperature Humidity µ top − µ bot (p-val) µ top − µ bot (p-val) 91 OH -0.004 (0.750) -0.007 (0.489) 101 OH 0.001 (0.909) 0.004 (0.457) 105 TX 0.041 ( < 0 . 000 ) -0.028 (0.001) 119 MN -0.008 (0.699) -0.013 (0.337) 123 TX 0.017 (0.002) 0.029 ( < 0 . 000 ) 127 NM 0.032 ( < 0 . 000 ) -0.063 ( < 0 . 000 ) 135 OH -0.045 (0.010) 0.017 (0.278) 144 CA 0.009 ( < 0 . 000 ) -0.018 (0.002) 145 CA -0.001 (0.675) 0.004 (0.549) 147 CA 0.011 ( < 0 . 000 ) -0.013 (0.017) 149 CA -0.007 (0.025) 0.008 (0.214) 153 CT 0.043 ( < 0 . 000 ) -0.003 (0.746) 155 NY 0.093 ( < 0 . 000 ) 0.012 (0.341) 156 NC 0.040 (0.007) -0.041 (0.445) 157 OH -0.132 ( < 0 . 000 ) -0.020 (0.638) 163 OH 0.180 (0.010) 0.179 (0.021) 168 P A 0.012 (0.122) 0.071 (0.006) 170 IL -0.001 (0.772) -0.007 (0.642) 173 OH 0.037 (0.003) -0.033 (0.440) T ABLE V: The difference in means and paired t-test p-value results, obtained by comparing temperature/humidity values among the bottom 10% and top 10% of hand-hygiene compli- ance rates, by facility ( blue indicates that either temperature, humidity , or both hav e a positi ve difference in means and a p-value ≤ . 05 ). T able V sho ws that most facilities have statistically signif- icant dif ferences between the two samples and that µ top 10 > µ bottom 10 . Such results indicates that higher temperatures and lev els of humidity (particularly temperature) are statistically associated with higher rates of hand hygiene. Ho wev er , we find that some facilities co-located in the same geographic region hav e conflicting statistical results (e.g., Facs. 91, 173). W e conjecture that such a result may attributable to differences in sensor deployment location, but we leave such an in vestigation as future work. E. F acility-Specific Modeling The full M 5 Ridge Regression models’ reliance on facility identities suggests that compliance relies, at least in part, on facility-specific health care worker attitudes, administrative culture, or e ven simply the disposition of sensors and the archi- tecture of the facility . Given the magnitude of the coefficients associated with facilities in the previous model, we propose to construct and analyze a facility-specific model. Here, we selected a facility (facility 91) with both a high rate of compliance and a large number of reported e vents for further in vestigation (see T able VI). As expected, the facility- specific model is better at predicting compliance than the full model (T able II), while the correlation is comparable. The hypothesis terms associated with this model are shown in T able VII. Unlike the previous model, temperature is (a) Facility 91. (b) Facility 101. (c) Flu Sev erity . (d) Humidity . (e) T emperature. Fig. 2: The marginal effects of sev eral select cov ariates, where blue shows the kernel density of the original data and the red lines show the estimation. Rate (y-axis) vs. feature (x-axis). Note that in 2a and 2b no kernel density estimate is provided, as these plots are for binary features. Measure V alue Correlation 0.3179 RMSE 0.0381 T ABLE VI: Correlation coefficient and RMSE of a cross- validated model for Facility 91. negati vely associated with compliance, which is some what surprising. W e also note a larger negati ve association between compliance and flu severity which, while somewhat harder to explain, may also reflect the narrower geographic scope accounted for by this model. Ultimately , only w eekday and humidity positi vely impact compliance, which is a dif ferent result than in our global model. These differences aren’t surprising, howe ver: the original model attempts to capture effects across a broad geographic region, while this model need only capture the associations found in a specific location. Featur e h j Air T emp − 0 . 0858 Rel. Humid. 0 . 0546 W eekday 0 . 039 Day Shift − 0 . 1742 Flu Sev . − 0 . 2097 T ABLE VII: Feature specific h j terms for the Facility 91 model, where red highlights features with a negati ve asso- ciation and blue highlights those with a positiv e association. I V . D I S C U S S I O N A N D F U T U R E W O R K In this section we discuss the broader implications of our findings, as well as directions for future work. The full model and marginal effects models, in conjunction with the RReliefF feature ranking, provided sev eral insights. First, we found that facility identities are strongly related to compliance, suggesting that facility-wide attitudes towards hand hygiene exist, persist in time, and are predicti ve of compliance rates. This observation may also reflect differences in sensor installation, where different facilities may hav e sensors instrumented in dif ferent departments, thus affecting reported rates. Second, increases in influenza se verity were associated with an increase in compliance, which is encour- aging. Third, our conjecture regarding lower weekend and holiday compliance appears to have some merit, although the holidays associated with negati ve compliance were somewhat surprising. W e again acknowledge that this result may be affected by increased visitors during these times. Fourth, our conjectures that higher humidity and temperature are indicati ve of higher rates of compliance were confirmed by the full model, marginal effects model, and statistical analysis. This finding is important as health care workers often cite skin irritation or dry skin as reasons for reduced frequency of hand hygiene. Fifth, we found that compliance during the first week of residents’ attendance ran contrary to our original conjecture: the J ul y E f f ect was essentially unobservable. Finally , we found that nig htS hif t was associated with slightly lower compliance rates. Our facility-specific model (constructed for Facility 91) found contradictions with the full hypothesis. W e believe that this supports the facility-specific attitudes conjecture and that, moreov er , dif ferent factors may be at play at different facilities spanning dif ferent geographical re gions. Further work is needed to tease these differences out, howe ver . This work has se veral limitations. First, there are dif ferences among installations: not all doors and dispensers may be instrumented and, therefore, we cannot track, for example, the use of personal alcohol dispensers (we assume stable practices). Thus our compliance estimates may be based on partial information and are certainly not comparable across facilities. Second, our compliance estimates are facility wide, meaning that we do not exploit the co-location of dispensers and door ev ent sensors, but only the temporal correlation of the individual events. Thus, our assumption that each door ev ent corresponds to a hand-hygiene opportunity may be fundamentally flawed, even as it allows for consistent intra- facility comparisons. Third, we acknowledge the possibility of location and sampling bias with regard to both the sensors and facilities. If sensors were to be placed in only the ICU of one facility and in the emergenc y room of another , we may observ e different rates, which has not been accounted for . Additionally , though facilities are distributed across the United States, they are by no means meant to be a representativ e sample of facility types or climatic conditions. There are also a number of opportunities for future work. First, we would like to consider alternati ve definitions of compliance and examine compliance at finer-grained tempo- ral levels, perhaps incorporating time-series analyses as an additional a venue of exploration. W e intend to also explore framing the problem as one of classification, rather than only regression, which may help tease out uncertain f actors. Finally , data pertaining to compliance rates under certain interventions would giv e way to exploration of intervention efficac y both in general and using prediction-based methodology , such as in verse classification, to recommend facility-specific interven- tion policies [28], [29]. Hand hygiene compliance is a simple yet effecti ve method of pre venting the transmission of disease, both among the population at large, and within health care facilities. This study presents a first look at factors that underlie health care worker hand-hygiene compliance rates, including weather conditions, holidays and weekends, and infectious disease prev alence and sev erity , and serves as a model for future studies that will exploit the av ailability of temporally and spatially rich compliance data collected by the sophisticated sensor systems now being put into practice. A C K N O W L E D G M E N T S The authors would like to thank Gojo Industries for access to the hand-hygiene data and for their financial support of this work, as well as Andrew Arthur for his help in preparing the data. R E F E R E N C E S [1] R. Kle vens, J. Edwards, C. Richards, and T . Horan, “Estimating health care-associated infections and deaths in us hospitals, ” Public Health , no. 122, pp. 160–166, 2007. [2] R. Roberts, R. Scott, B. Hota, L. Kampe, F . Abbasi, S. Schabowski, I. Ahmad, G. Ciavarella, R. Cordell, S. Solomon, R. Hagtvedt, and R. W einstein, “Costs attributable to healthcare-acquired infection in hospitalized adults and a comparison of economic methods, ” Medical Car e , vol. 48, no. 11, pp. 1026–1035, Nov ember 2010. [3] R. Roberts, B. Hota, I. Ahmad, R. Scott, S. Foster , F . Abbasi, S. Sch- abowski, L. Kampe, G. Ciavarella, M. Supino, J. Naples, R. Cordell, S. Levy , and R. W einstein, “Hospital and societal costs of antimicrobial- resistant infection in a chiago teaching hospital: implications for an- tibiotic stewardship, ” Clinical Infectious Diseases , vol. 49, no. 8, pp. 1175–1184, October 2009. [4] J. M. Boyce and D. Pittet, “Guidelines for hand hygiene in health-care settings: recommendations of the healthcare infection control practices advisory committee and the hicpac/shea/apic/idsa hand hygiene task force, ” Infection Contr ol and Hospital Epidemiology , no. 23, pp. S3– S41, 2002. [5] B. Allegranzi, H. Sax, L. Bengaly , H. Richet, D. Minta, M. Chraiti, F . Sokona, A. Gayet-Ageron, P . Bonnabry , and D. Pittet, “W orld health organization ”point g” project management committee. successful imple- mentation of the world health organization hand hygiene improvement strategy in a referral hospital in mali, africa, ” Infection Contr ol and Hospital Epidemiology , vol. 31, no. 2, pp. 133–141, February 2010. [6] D. Pittet, B. Allegranzi, and J. Boyce, “W orld health organization world alliance for patient safety first global patient safety challenge core group of experts. the world health organization guidelines on hand hygiene in health care and their consensus recommendations, ” Infection Control and Hospital Epidemiology , vol. 30, no. 7, pp. 611–622, July 2009. [7] J. P . Hass and L. E. L., “Measurement of compliance with hand hygiene, ” Journal of Hospital Infection , no. 66, pp. 6–14, 2007. [8] J. Boyce and M. Cooper , T anda Dolan, “Evaluation of an electronic device for real-time measurement of alcohol-based hand rub use, ” Infection Control and Hospital Epidemiology , v ol. 30, no. 11, pp. 1090– 1095, 2009. [9] Joint Commission of Accreditation of Healthcare Organizations, “Patient safety goals, ” T ech. Rep. [Online]. A vailable: http://www .jcaho.org/accredited+or ganizations/patient+safety/npsg.htm [10] J. Fries, A. Segre, G. Thomas, T . Herman, K. Ellingson, and P . Polgreen, “Monitoring hand hygiene via human observers: How should we be sampling?” Infection Control and Hospital Epidemiolo gy , v ol. 33, no. 7, pp. 689–695, Jul. 2012, [PMID: 22669230]. [11] D. Sharma, G. Thomas, E. Foster , J. Iacovelli, K. Lea, J. Streit, and P . Polgreen, “The precision of human-generated hand-hygiene observ a- tions: a comparison of human observation with an automated monitoring system, ” Infection Contr ol and Hospital Epidemiology , vol. 33, no. 12, pp. 1259–1261, December 2012. [12] T . Eckmanns, J. Bessert, M. Behnke, and H. Gastmeier, P anda Ruden, “Compliance with antiseptic hand rub use in intensive care units: The hawthorne effect, ” Infection Control and Hospital Epidemiology , no. 27, pp. 931–934, 2006. [13] M. Monsalve, S. Pemmaraju, G. Thomas, T . Herman, and P . Segre, AM anda Polgreen, “Do peer effects improve hand hygiene adherence among healthcare workers?” Infection Contr ol and Hospital Epidemiology , vol. 35, no. 10, pp. 1277–1285, 2014. [14] V . Boscart, K. McGilton, A. Levchenko, G. Hufton, P . Holliday , and G. Fernie, “ Acceptability of a wearable hand hygiene device with monitoring capabilities, ” Journal of Hospital Infection , vol. 70, no. 3, pp. 216–222, 2008. [15] A. V enkatesh, M. Lankford, D. Rooney , T . Blachford, C. W atts, and G. Noskin, “Use of electronic alerts to enhance hand hygiene compliance and decrease transmission of vancomycin-resistant enterococcus in a hematology unit, ” vol. 36, no. 3, pp. 199–205, 2008. [16] P . M. Polgreen, C. S. Hlady , M. a. Severson, A. M. Segre, and T . Herman, “Method for automated monitoring of hand hygiene ad- herence without radio-frequency identification.” Infection contr ol and hospital epidemiology : the official journal of the Society of Hospital Epidemiologists of America , vol. 31, no. 12, pp. 1294–1297, 2010. [17] H. Dai, K. L. Milkman, D. A. Hofmann, and B. R. Staats, “The Impact of T ime at W ork and T ime Off from W ork on Rule Compliance: The Case of Hand Hygiene in Healthcare, ” Journal of Applied Psychology , vol. 100, no. 3, pp. 846–862, 2014. [Online]. A vailable: http://papers.ssrn.com/sol3/papers.cfm?abstract id=2423009 [18] C. Jarrin T ejada and G. Bearman, “Hand Hygiene Compliance Monitoring: the State of the Art, ” Curr ent Infectious Disease Reports , vol. 17, no. 4, 2015. [Online]. A vailable: http://link.springer .com/10.1007/s11908-015-0470-0 [19] E. Kalnay , M. Kanamitsu, R. Kistler, W . Collins, D. Dea ven, L. Gandin, S. Iredell, S. Saha, G. White, Y . Zhu, a. Leetmaa, R. Reynolds, M. Chelliah, W . Ebisuzaki, W . Higgins, J. Janowiak, K. Mo, C. Ropelewski, J. W ang, R. Jenne, and D. Joseph, “The NCEP/NCAR 40-Y ear Reanalysis Project, ” pp. 437–471, 1996. [Online]. A vailable: [20] N. R. Draper , H. Smith, and E. Pownell, Applied Regr ession Analysis . W iley New Y ork, 1966, vol. 3. [21] J. R. Quinlan, “Learning with continuous classes, ” in 5th Australian Joint Confer ence on Artificial Intelligence , vol. 92, 1992, pp. 343–348. [22] F . D. Johansson, U. Shalit, and D. Sontag, “Learning representations for counterfactual inference, ” in 33r d International Confer ence on Machine Learning (ICML) , 2016. [23] R. Tibshirani, “Regression shrinkage and selection via the lasso, ” Journal of the Royal Statistical Society . Series B (Methodological) , pp. 267–288, 1996. [24] H. Zou and T . Hastie, “Regularization and variable selection via the elastic net, ” Journal of the Royal Statistical Society: Series B (Statistical Methodology) , vol. 67, no. 2, pp. 301–320, 2005. [25] M. Robnik- ˇ Sikonja and I. Kononenko, “ An adaptation of relief for attribute estimation in regression, ” in Machine Learning: Proceedings of the F ourteenth International Conference (ICML97) , 1997, pp. 296– 304. [26] K. Kira and L. A. Rendell, “ A practical approach to feature selection, ” in Pr oceedings of the ninth international workshop on Machine learning , 1992, pp. 249–256. [27] R. Williams et al. , “Using the margins command to estimate and interpret adjusted predictions and marginal effects, ” The Stata Journal , vol. 12, no. 2, p. 308, 2012. [28] M. T . Lash, Q. Lin, W . N. Street, J. G. Robinson, and J. Ohlmann, “Generalized in verse classification, ” in Proceedings of the 2017 SIAM International Conference on Data Mining (SDM’17) . [29] M. T . Lash and W . N. Street, “Realistic risk-mitigating recommendations via in verse classification, ” arXiv pr eprint; arxiv:1611.04199 , 2016. [Online]. A vailable: https://arxiv .org/abs/1611.04199