Modelling Temporal Relationships in Pseudomonas Aeruginosa Antimicrobial Resistance Prediction in Intensive Care Unit Àlvar Hernàndez-Carnerero 1 and Miquel Sànchez-Marrè 2 and Inmaculada Mora-Jiménez 3 and Cristina Soguero-Ruiz 4 and Sergio Martı́nez-Agüero 5 and Joaquı́n Álvarez-Rodrı́guez 6 Abstract. In this paper, the prediction of antimicrobial resistance distorting the anatomical integrity-protective barriers of patients (in- of Pseudomonas aeruginosa bacteria caused by nosocomial infec- tubation, mechanical ventilation, vascular access, etc.) [3]. tions in the Intensive Care Unit (ICU) was considered. It was trained It is frequent to find in the ICU some kinds of bacteria which can a Logistic Regression model using health records information from become multidrug-resistant. Among them, Acinetobacter spp., Ente- patients together with the history of past sensitivity tests (antibi- rococcous fecalis and Enterococcus faecium, Escherichia coli, Kleb- ograms). To predict the antimicrobial resistance for a certain patient, siella pneumoniae, Pseudomonas aeruginosa and Staphylococcus this study proposes to model the temporal relationships using bacte- aureus are usually found. In this paper, we focus on Pseudomonas rial information from the rest of the patients who are at the same time aeruginosa due to their prevalence and virulence. It is naturally resis- in the ICU. Furthermore, a training window with incremental size is tant to many antibiotics and has a remarkable capacity for acquiring used so that training set is always temporarily as near as possible to new resistance mechanisms, creating therapeutic problems [6]. Pseu- test instances to be predicted. Using these contributions, experiments domonas aeruginosa is considered to be multi-drug resistant (MDR) show promising results to predict antimicrobial resistance even when when it is observed a reduced in vitro susceptibility to three or more few training data is available. From these results it is further inferred antimicrobial families [21]. that resistant bacteria may be spreading among patients in the ICU It is known that infections due to MDR microorganisms are a ma- and their populations rapidly mutate, changing the underlying data jor problem. This has a significant impact in the ICU, where they can distribution, along time. cause additional morbidity, mortality, and health care costs [3, 11]. Inappropriate initial antimicrobial treatment of P. aeruginosa is sta- tistically linked to a higher mortality compared to initial treatment 1 INTRODUCTION with appropriate antimicrobial. The growing MDR rate of P. aerug- Antimicrobial resistance has been increasing for decades, and the inosa also increases the chance of inappropriate initial antimicrobial rate at which new antibiotics are synthesized is not as fast as it would treatment [13]. be required to prevent this trend [7, 17]. A large proportion of in- To tackle with MDR in the hospital, a culture or microbiological fections caused by resistant bacteria occur during hospital stays, spe- analysis is usually performed to test whether the bacterium is resis- cially in the Intensive Care Unit (ICU) [16]. There, infection rates are tant or susceptible to a set of antibiotics. For this purpose, first the much higher than in other hospital divisions [9]. This is due to its germ (bacterium) is isolated and an antibiogram is carried out. The severely vulnerable population and to the high risk of becoming in- antibiogram represents the in vitro bacterium’s resistance to a series fected through multiple procedures and to the use of invasive devices of antibiotics. The set of antibiotics used in the antibiogram can be 1 Campus del Poblenou, Universitat Pompeu Fabra, 08018 Barcelona, Spain, selected for the specific type of bacteria being tested. The result of email: alvar.hernandez01@estudiant.upf.edu the antibiogram is a vector of pairs antibiotic/sensibility [4]. Antibi- 2 Knowledge Engineering and Machine Learning Group (KEMLG-UPC), In- ograms are often used by clinicians to assess bacteria susceptibility telligent Data Science and Artificial Intelligence Research Centre (IDEAI- rates, as an aid in selecting empiric antibiotic therapy [10]. Hence, the UPC), Dept. of Computer Science, Universitat Politècnica de Catalunya, antibiogram result could vary between bacterium species, depending Spain, email: miquel@cs.upc.edu 3 Department of Signal Theory and Communications, Telematics and Com- on the resistance of a particular bacterium to different antibiotics. puting Systems, Rey Juan Carlos University, Madrid 28943, Spain, email: However, quite often groups of antibiotics still have similar sensi- inmaculada.mora@urjc.es tivity when tested on a given bacterium species, despite its strains 4 Department of Signal Theory and Communications, Telematics and Com- [20]. puting Systems, Rey Juan Carlos University, Madrid 28943, Spain, email: cristina.soguero@urjc.es In addition, the result of the antibiogram can help to reduce the 5 Department of Signal Theory and Communications, Telematics and Com- bacteria spread by taking special measurements such as isolation of puting Systems, Rey Juan Carlos University, Madrid 28943, Spain, email: the patient. One of the most relevant factors of the spread of bac- sergio.martinez@urjc.es terial resistance is the so called cross-transmission [18], which may 6 Intensive Care Department, University Hospital of Fuenlabrada, Madrid 28942, Spain, email: joaquin.alvarez@salud.madrid.org facilitate the spread of resistant bacteria from one patient to another. Copyright © 2020 for this paper by its authors. Use permitted under Cre- Also, some measures such as hand hygiene, skin cleansing, and con- ative Commons License Attribution 4.0 International (CC BY 4.0). This tact precautions can help to prevent cross-transmission [18]. To know volume is published and copyrighted by its editors. Advances in Artificial how and where to extreme caution, information about the kind of Intelligence for Healthcare, September 4, 2020, Virtual Workshop. bacteria in the ICU and their resistance plays a key role. Since ICU patients have a critical health status and the antibiogram result can take from 24h to 48h [12], it is of major importance to develop tools which can help to anticipate this result. This would contribute not only to save patient’s lives but also to prevent the spread of a resis- tant bacteria. Because of the aforementioned reasons, in the current study we Table 1. Feature names and their description. Marked in bold the target propose to use a Data Mining (DM) technique, and specific tempo- features. ral data processing to get a quick estimation of the antibiogram re- Feature name Description sult. In this sense, many of the state-of-the-art studies regarding the use of DM methods to predict antimicrobial susceptibility are using Result of the sensibility test to the c&car whole genome sequencing [8, 15, 14, 1]. Despite it is a very promis- Carbapenems family (r/s). Result of the sensibility test to the ing technique, it involves very significant costs. As an alternative, we c&cf4 4th Generation Cephalosporins propose to use information from the ICU health records and demo- family (r/s). graphic data of patients, along with historic antibiogram results to days to culture Number of days elapsed from train a DM model, aiming to predict resistant bacteria in new cul- admission to the date of the culture. date culture Date of the culture. tures. As opposed to the whole genome sequencing, our approach Number of antibiotics tested in the intends to use data which are already available in the vast majority number antibiotics antibiogram. of hospitals, in order to speed up the process of identifying positive Type of culture performed (pharynx, culture type cases. Similar strategies have been analyzed in the past [12, 20, 19]. urine, blood, etc.). The remainder of this paper is structured as follows. Section 2 de- Type of culture performed grouped culture type grouped (respiratory, urine, surface, etc.). scribes the dataset and the procedure to create new features. The ex- Type of culture performed grouped perimental setup is established in Section 3, and results in Section 4. culture type grouped 2 (clinical sample/surface). Finally, conclusions, limitations of the study, and suggested future day month culture Day of the month on which research are presented in Section 5. culture is carried out. Month on which culture is month culture carried out. Year on which culture is year culture carried out. 2 DATA DESCRIPTION AND PREPROCESSING Clinical origin before ICU origin admission. reason admission Reason of admission at ICU. Group of illness A. Considers goi A Data considered in this work is a unified and anonymized dataset cardiovascular events. specifically collected for the study of antimicrobial resistance in the Group of illness B. Considers goi B kidney failure, arthritis. ICU of the University Hospital of Fuenlabrada (UHF) in Spain. The Group of illness C. Considers data set covers the years from 2004 to 2013. During this time interval, goi C respiratory problems. 1914 patients were admitted to the ICU, and 22142 cultures were Group of illness D. Considers goi D carried out from 2186 admissions. It has a number of 257 different pancreatitis, endocrine. Group of illness E. Considers types of bacteria and 26 antimicrobial families. goi E epilepsy, dementia. The dataset contains the results of antibiograms carried out to pa- Group of illness F. Considers goi F tients in the ICU, that is, the results of the sensitivity tests (suscep- diabetes, arteriosclerosis. tible (s); or resistant (r)) for certain pairs of bacteria and family of Group of illness G. Considers goi G antibiotic used in the test. It also includes demographic data of the neoplasms. Number of groups of illness to patients and information of their ICU admission: age, gender, date pluripathology which patient belongs. of ICU admission, clinical origin of the patient before ICU admis- patient category Patient’s clinical category. sion, reason for admission, patient category, comorbidities and pluri- age Patient age. pathology (it indicates whether a patient has more than two comor- gender Patient’s gender. Date on which the patient’s bidities). start date admission begins. As already mentioned, in this study we focused on just one type Day of the week on which the day week admission of bacteria among the multiple available in the data set: P. aerugi- patient’s admission begins. nosa. This bacterium is considered MDR if it is resistant to three or Day of the month on which the day month admission more of the following antimicrobial families within the same culture: patient’s admission begins. Month on which the patient’s Aminoglycosides, Carbapenems, 4th Generation Cephalosporins, month admission admission begins. Extended-spectrum penicillins, Polymyxins and Quinolones. Year on which the patient’s year admission In this work, we analyze all instances containing the bacterium admission begins. and antimicrobial families of interest. Then, each instance is repre- sented by features described in Table 1. Features c&car and c&cf4 represent the target to be predicted, this is, the result of the sensibility test for P. aeruginosa to the antimicrobial families of Carbapenems and 4th Generation Cephalosporins, respectively. We consider only these antimicrobial families since this is the first approach to analyze the problem and allows us to reduce the scope of this study. 2.1 Generation of new features Table 2. Features with missing values (expressed in %) in both data sets. In addition to the selected features, we propose to generate a new kind of features based on the temporal information of cultures Dataset r&amg r&car r&cf4 r&pap r&pol r&qui recorded in the data set. c&car 11.33 15.56 11.33 11.33 39.33 11.78 The purpose of these features is to capture the presence of resistant c&cf4 10.79 17.99 10.79 10.79 35.97 11.15 bacteria in the ICU along time and the ”intensity” of that presence. By ”intensity”, we consider the number of patients infected and the number of days since resistant bacterium was detected. For a specific aeruginosa and a particular antimicrobial family for some time inter- instance of the data set, which represents a culture Cp of patient Pp , vals, and therefore, a missing value is considered for that feature.This cultures containing P. aeruginosa have been collected for patients Pi fact can be addressed following different approaches, such as delet- (Ci = {Cij }) between 21 days and 48 hours before the date of the ing instances with missing data or imputing missing values [22]. In culture Cp . These cultures exclude those associated to the patient Pp this study we propose an strategy based on the clinical meaning of under analysis. Note that, since the results of the test usually takes the generated features. The smaller the value of these features, the 48 hours to be provided, it is not possible to use cultures taken, for fewer patients will have been infected with resistant P. aeruginosa instance, one hour ago. Apart from that, from a clinical viewpoint, bacteria and the greater the length of time since they were infected. if the culture result is positive, it is kept as positive for the next 21 As a result, if r&* features do not have any value, it suggests that no days. For this reason, we consider cultures collected 21 days before P. aeruginosa has been recently detected in the ICU. Therefore, very the date Dp of the current culture Cp of patient Pp . likely no patients would have been recently infected with a resistant In total, six features are created using the information of the past bacteria, and the value provided by Eq. (2) should be very small. In cultures, one for each type of the antimicrobial families mentioned this case, missing values are replaced by a 0. above: r&amg, r&car, r&cf4, r&pap, r&pol and r&qui. For Afterwards, all categorical features are converted to binary follow- instance, feature r&amg identifies only cultures tested for the Amino- ing a one-hot encoding strategy, except for the features representing glycosides family. The value of this feature is obtained taking into dates. Since dates have an intrinsic ordering, smaller numerical val- account the set of past cultures for all other patients (Pi ), and ex- ues are assigned to further dates in the past, and greater values corre- cluding the patient under analysis (Pp ). Each culture Cij on date spond to more recent dates. dij for patient Pi has the sensibility test result rij , which is 0 or 1 Finally, Pearson correlation between features (without consider- depending on whether the bacterium is susceptible or resistant to a ing the targets) is calculated in order to discard the most correlated specific family of antibiotics. In order to give more emphasis to the features, since they provide very similar information. The method- most recent cultures, we use a negative exponential function [2] to ology is as follows. When two features have a correlation coeffi- weight the culture results associated to each patient Pi as follows: cient higher than 0.9 or lower than -0.9, just one of them is ran- ( domly selected, discarding the rest. A visual representation of Pear- 0 if rij = 0 son correlation between features for Carbapenems and 4th Genera- fCij (Dp ) = (1) n−(Dp −dij ) if rij = 1 tion Cephalosporins subsets is shown in Fig. 1. We can conclude that a similar correlation in patterns is found for both subsets. In where, n is a real number experimentally set to 1.1. Then, to compute both, there is just one group of features that are correlated more than the value of each of the six features linked to patient Pp on date Dp , 0.9, which are the following: date culture, year culture, for each patient excepting Pp the maximum outcome in Eq. (1) is start date and year admission. From these four correlated determined and added up according to Eq. (2). features, date culture is selected because it is the most repre- X sentative among them and the rest are discarded. Finally, both data F VCp (Dp ) = max fCij (Dp ) (2) Ci sets had 31 features after this discarding. ∀Pi 6=Pp 2.2 Data preprocessing To proceed with the model design for c&car and c&cf4, we created two data subsets, one associated to each target. These subsets will be used to train a Logistic Regression (LR) model for each target. Train- ing two different classifiers instead of, for instance a multi-class clas- sifier, allows each classifier to be specialized in predicting its partic- ular target, therefore tuning classifier’s weights individually. In order to limit the study to MDR adquired in the ICU, only instances of patients admitted in the ICU for more than 48h are considered. Figure 1. Visual representation of Pearson correlation between features The final data set for the c&car was composed by 450 cultures for: (a) Carbapenems and (b) 4th Generation Cephalosporins subset. and 34 features including the target one, and the final data set for the r&cf4 was composed by 556 cultures and the same 34 features in- cluding the target one. In both data sets there are only missing values in six features. The percentages of missing values in those features 3 EXPERIMENTAL SETUP are depicted in the Table 2. Before training the models, we deal with missing data associated The first experiment, once data is processed, is to evaluate the rele- to the proposed features r&amg, r&car, r&cf4, r&pap, r&pol vance of the set of features selected and proposed, regarding the two and r&qui. Note that there may not be any culture in the ICU for P. different target features to be predicted. This is going to be analyzed by using Mutual Information (MI) [5]. It is a quantity that measures feasibility of learning the data. The classifier has to decide whether the mutual dependence of the two variables, that is, it quantifies the the target is sensitive or resistant. Therefore it is a binary decision. amount of information that a random variable provides about other To assess the performance of the proposed incremental window random variable. In terms of the probabilities, the MI of two jointly framework, experiments are carried out with different configurations. discrete random variables X and Y is calculated as: The characteristics of defined training and test windows are the fol- lowing:   XX p(X,Y ) (x, y) • For each experiment a set of LR classifiers is trained, each for a I(X; Y ) = p(X,Y ) (x, y) log (3) y∈Y x∈X pX (x) pY (y) different training-test window. • The size chosen for test windows is fixed in 3 months. Which is where p(X,Y ) is the joint probability mass function of X and Y , a relatively short time, near to the training instances, containing and pX and pY are the marginal probability mass functions of X and enough test instances. Y respectively. • Different test windows do not overlap between them. That is, com- Looking now at the prediction of the target, the type of the problem pared to the others, each test window contains different instances proposed have associated a series of special characteristics that have belonging to different time intervals. to be considered in order to properly address it. • Instances in the training window do not contain antibiograms be- First property in health records have an inherent temporal order- longing to patients that also are present in the test window. For ing. This forces, to use as training only instances that belong to a time instance, if the result of an antibiogram of a particular patient is prior to the test antibiograms that are to be predicted. In addition a to be predicted in the test set, there are not past antibiograms of margin of time has to be respected between train and test windows, the same patient in the training set. That way, it is ensured that since results of antibiograms are not immediately available after they patients from training and test windows are different. are carried out. As before, in this particular case, a time margin of 48h needs to be considered. With each training and test pair of windows, a simple validation is The second particularity encountered when trying to predict is the performed to maintain the temporal order. The class imbalance due concept drift. It is the fact that the concept of interest may depend on to the nature of the problem, causes that in the test window there some hidden context, not given explicitly in the form of predictive might be more instances of one class than the other. To get a re- features. Changes in the hidden context over time can induce more alistic approximation of the algorithm performance, the true nega- or less radical changes in the target concept. Changes in hidden con- tives (success in sensitive instances) and the true positives (success text may not only result in a change of the target concept, but may in resistant instances) are calculated, together with the average be- also cause a change of the underlying data distribution [20]. In the tween these two values and the general accuracy. For a particular test particular domain of this study, antimicrobial resistance, the hidden window with ns sensitive instances and nr resistant instances, if the context that changes over time are the mutations of bacteria, that al- method succeeds in predicting ps sensitive instances and pr resistant low them to be more resistant to antibiotics, as time passes by. In [20] instances, the just mentioned values are calculated as: it is proposed to use instance selection to handle concept drift as it ps + pr is the technique that is most commonly used and has been found to general accuracy= (4) ns + nr offer good results. More specifically it is proposed to use a technique pr based on instance selection that consists in generalizing from a win- resistant success= (5) dow that moves over recently arrived instances and uses the learnt sr ps concepts for prediction only in the immediate future. This represents sensitive success= (6) a very good approach to apply to the resolution of the problem ana- ns pr lyzed in this study except for the third particularity of the data set. s + npss average success= r (7) The data scarcity makes it difficult to learn from temporal win- 2 dows containing several months, even years. In the first data set con- These metrics offer a better approximation of the performance, be- sidered, the one with target feature c&car there are a total of 450 in- cause allow to track the success rate in the minority class label, which stances, and for the second data set with target feature c&cf4 there in many real problems is the most important one. For instance, if are 556. Taking into account those data sets represent cultures from the test set counts with 8 sensitive and 2 resistant instances and the 10 years (from 2004 to 2013), the average of cultures per year is 45 DM technique predicts all instances as sensitive, the general accuracy and 56 respectively, which is a relatively small number of instances metric would pretty high value of a 80%, while it will be perform- considering how fast ICU bacteria is able to mutate and change its ing poorly in indentifying resistant antibiograms which are the ones sensitivity patterns. For this reason, we propose to use an incremental most needed to detect. window for training, which increases its size as test window moves Finally, to calculate the mean accuracy among several windows, towards more recent instances. That is, the training window will be an accumulation of the success rates is done, and the performance is fixed from the first temporal instances, which are the oldest, and it evaluated using Equations 4, 5, 6 and 7. In other words, the perfor- will gradually increment in size, containing more instances, as more mance of several windows is not calculated by averaging the indi- recent instances of test are predicted. The test window, on the other vidual performance values, but by accumulating the number of suc- hand, will have a fixed size and it will progressively slide to select cess instances in each window and using it to calculate the accuracy. more recent instances. This is done because test windows may have a different number of The DM technique used in experiments is Logistic Regression instances, due to the fact that not all 3-month time intervals have (LR). It is chosen because of its simplicity to be used as a baseline. A the same number of antibiograms. Therefore, making an average be- baseline is a model that is both simple to set up and has a reasonable tween their accuracy values would not be adequate since some in- chance of providing acceptable results. LR is a technique which is stances would have more weight than others depending on the num- used for the classification, and is used in this study to evaluate the ber of instances in their test window. 4 RESULTS This section presents the results obtained after carrying out the exper- iments described in Section 3. In Section 4.1, the mutual dependence among features is analyzed by using MI. Sections 4.2 and 4.3 eval- uate the impact of features, date culture and r&* respectively, on the prediction of bacteria resistance. The improvement in predic- tion obtained by using the incremental training window scheme is assessed in Section 4.4. 4.1 Features relevance using mutual information The results of feature relevance using MI for Carbapenems and 4th Table 3. MI feature weighting for c&car and c&cf4 target. Generation Cephalosporins targets are represented in left and right columns of Table 3 respectively. The method MI computes the rel- evance or weight of one feature according to the co-occurrence of Feat. name Wgts. Feat. name Wgts. c&car c&cf4 this feature and the target feature as described in equation 3. The MI method does not take into account the possible interaction of date culture 0.5312 date culture 0.4440 this feature with other ones regarding the target feature. In both age 0.1866 r&qui 0.1637 days to culture 0.1713 age 0.1578 cases, the feature date culture is by far the variable containing r&qui 0.1711 days to culture 0.1505 most information about the feature to be predicted. Regarding Car- r&pap 0.1668 r&pap 0.1491 bapenems, date culture has a value of 0.53, while the second r&amg 0.1459 r&amg 0.1295 most important feature which is age just receives a relevance value r&car 0.1382 r&car 0.1289 of 0.18. In 4th Generation Cephalosporins date culture gets a r&cf4 0.1228 r&cf4 0.1137 day month- day month- value of 0.44 and r&qui, the second one 0.16. The importance of 0.1167 0.1038 admission admission date culture suggests that results of antibiograms are highly de- origin 0.0981 number antibiotics 0.0799 pendent on the time they were performed. number antibiotics 0.0936 origin 0.0792 In addition, it is notable that five of the six proposed features r&*, reason admission 0.0860 reason admission 0.0608 day month culture 0.0509 month admission 0.0556 which consider antibiograms of other patients in the ICU, are be- month culture 0.0333 goi E 0.0485 tween the eight most relevant features for both Carbapenems and culture type 0.0279 day month culture 0.0393 4th Generation Cephalosporins. Therefore, we can infer they con- month admission 0.0219 day week admission 0.0366 tain a great amount of information to predict antimicrobial resis- day week admission 0.0214 culture type 0.0243 tance. The proposed feature with a considerably lowest importance culture type- 0.0190 r&pol 0.0209 grouped is r&pol. This is probably caused because in the data set there are r&pol 0.0144 pluripathology 0.0204 a smaller number of antibiograms for Polymyxins antimicrobial fam- pluripathology 0.0074 month culture 0.0131 ily (296) compared to the amount for Aminoglycosides (564), Car- culture type- gender 0.0061 0.0116 bapenems (450), 4th Generation Cephalosporins (556), Extended- grouped goi B 0.0040 patient category 0.0099 spectrum penicillins (558) and Quinolones (558). goi F 0.0025 gender 0.0082 Hence, one can infer that the result of the antibiogram for a partic- goi A 0.0020 goi B 0.0030 ular patient is dependent on the past results of antibiograms of other culture type- 0.0017 goi F 0.0026 patients in the ICU. To explain this fact we propose as a hypothe- grouped 2 sis that bacteria has been spreading from one patient to another in goi E 0.0008 goi C 0.0023 goi D 0.0003 goi G 0.0022 the ICU by cross-transmission. Therefore, the fact that a patient is culture type- infected with a resistant bacteria may increase the odds of another patient category 4.5e-05 0.0010 grouped 2 patient becoming infected as well. goi G 2.0e-05 goi A 6.1e-05 goi C 2.9e-06 goi D 4.1e-09 Table showing the weight given by MI to each of the features in 4.2 Prediction contribution of date culture descending order regarding its value. Left column contains values for feature Carbapenems antimicrobial family and right column for 4th Gen. Cephalosporins family. After observing that date culture is the most important feature according to MI, the behavior of this significant variable is evaluated when predicting the result of antibiograms. To do that the target fea- ture is predicted in two modes, one considering all features includ- ing date culture, and a second one discarding date culture from the features set. These predictions are made by using an incre- mental window for training and a 3-months sliding window for test as described in the experimental setup section. The training window starts containing 2004 and 2005 instances and the test window the first three months of 2006. After that, the training window increases three months and test window slides three months until the end of the database is reached. Results of the accumulated accuracy for all win- dows are shown in the two left columns of Table 4 for Carbapenems and in the two right columns of the same Table 4 for 4th Generation Table 5. Year 2006 to 2013 prediction for Carbapenems and 4th Gen. Cephalosporins. Cephalosporins with incremental window. It is remarkable that in both cases the success percentage slightly increases when date culture is not used. We state that this may Carbapenems 4th Gen. Ceph. be due precisely to the high influence this feature has on prediction Metric With Without With Without of antimicrobial resistance. For some time intervals where the ma- Accuracy 57.22 57.97 60.88 60.68 jority of instances belong to a particular label, this feature might be Resistant success 50.92 47.25 40.72 30.41 forcing the DM method to learn that, in this particular interval of Sensitive success 64.97 71.19 73.62 79.80 values of date culture, it is highly probable that the predicted Average 57.94 59.22 57.17 55.11 instance belongs to the majority label. That is, it may be introducing Comparison with and without using the set of r&* features for some kind of bias towards the majority label of the time interval. This prediction of years 2006 to 2013 for both c&car and c&cf4 target features. is reflected in the way resistant success and sensitive success varies when date culture is removed. For Carbapenems target it can be seen that when using this feature, success in resistant instances 4.4 Prediction contribution of incremental window increases and success in sensitive instances decreases. In Carbapen- scheme ems data set there is a higher number of resistant instances (238) than sensitive instances (212). This seems to indicate that the ma- At last, the usefulness of the incremental training widow and 3- jority class enhances its accuracy when using date culture, and month test window scheme is evaluated to assess whether it improves the minority class makes it worse. The same is observed for 4th gen- success metrics with respect to using training windows furthest from eration Cephalosporins.In 4th generation Cephalosporins columns, the test set, and the same 3-month test windows. In this experiment the opposite situation happens, when date culture is considered, years 2012 and 2013 are predicted. In Table 6 success metrics for Resistant success decreases and Sensitive success increases, and now Carbapenems data sets are calculated, for 2012 and 2013. Figure 2 the majority class are sensitive instances (350) as opposed to resistant shows the number of test instances in each 3-month test window dur- instances (206). This apparently shows the bias date culture ing the mentioned years. In the two left columns of Table 6 accuracy feature introduces on classifing instances as the majority class, be- improves from a 59%, using a fixed window of training from 2004 to cause of its high influence over the prediction. 2011, to a 68% when using the incremental window to predict 2012 In next experiments, date culture is discarded since it slightly instances. In the three columns on the right of Table 6 it is observed reduces the accuracy of the algorithm. that, when predicting 2013, accuracy raises from a 64% when us- ing training data until 2011 to a 71% when considering year 2012 as training too. When using the incremental window, the accuracy Table 4. Year 2006 to 2013 prediction for Carbapenems and 4th Gen. Cephalosporins with incremental window. remains the same in a 71%. The reason why, in this case using the incremental window does not increase the accuracy with respect to Carbapenems 4th Gen. Ceph. training data until 2012 may be that the number of instances from Metric With Without With Without first months of 2013 is small as it can be seen in Figure 2, and in- Accuracy 56.96 57.22 57.49 60.88 cluding them in the incremental training window may not increase Resistant success 56.88 50.92 25.26 40.72 the knowledge of the problem. Sensitive success 57.06 64.97 77.85 73.62 Average 56.97 57.94 51.55 57.17 Table 6. Year 2012 and year 2013 prediction for Carbapenems. Comparison with and without using date culture feature for prediction of years 2006 to 2013 for both c&car and c&cf4 target features. 2012 2013 Metric 04-11 Inc. 04-11 04-12 Inc. Accuracy 59.09 68.18 64.29 71.43 71.43 Resistant success 57.14 66.67 57.14 85.71 85.71 Sensitive success 100.0 100.0 71.43 57.14 57.14 4.3 Prediction contribution of r&* features Average 78.57 83.33 64.29 71.43 71.43 In the two left columns, comparison of the year 2012 prediction A similar experiment is carried out for proposed r&* features. Ac- using a fixed training window with instances from 2004 to 2011 and curacy metrics are calculated with and without them, for test years an incremental training window. In the three columns on the right, comparison of 2013 prediction using a fixed training window with between 2006 and 2013. In Table 5 the results of this experiment instances from 2004 to 2011, a fixed training window from 2004 to are represented. The accuracy remains almost the same whether r&* 2012 and an incremental training window. The target feature predicted features are used to predict or not, for both antimicrobial families. in all cases is c&car. The difference is observed in resistant success and sensitive success. When r&* features are taken into account, these two metrics are more balanced, which means that both the majority class and mi- Table 7 and Figure 3, represent the results of the same experiment nority class get a similar success rate, which is a desirable effect. for 4th Generation Cephalosporins data. In the two left columns of Moreover, one can note that in both antimicrobial families, resistant Table 7 accuracy for predicting 2012 is the same considering training success increases, which means that cultures from other patients are instances until 2011 and using the incremental window. As before, helping to better discriminate the resistant instances. Sensitive suc- that can be due to instances of first months of 2012 not providing cess decreases which is probably caused by the LR decision bound- enough knowledge to improve accuracy of prediction, although in ary, which after moving to better predict resistant instances is lower- this case there is a greater number of them as it is shown in Figure 3. ing its performance in recognizing sensitive instances. The three columns on the right of Table 7show that accuracy to pre- Figure 2. Number of test instances for Carbapenems in years 2012 to Figure 3. Number of test instances for 4th Gen. Cephalosporins in years 2013 with 3 months granularity. 2012 to 2013 with 3 months granularity. dict 2013 is maintained in a 71% when making use of training data formation to predict bacteria resistance in other patients, which could until 2011 and training data until 2012, which as we have just seen indicate that bacteria is spreading among ICU patients. It has also may be caused by 2012 instances not providing much information. been observed that features providing the specific temporal ordering Considering the incremental window for training, the percentage in- between all instances in the data set tend to decrease prediction accu- creases to an 80%. racy. Lastly, experiments indicate that using an incremental window These last experiments reveal that using training data temporarily for training, maintain success rates or improve them. Therefore, we as close as possible to the test set, always maintain the same suc- can conclude it is a scheme that improves the algorithm performance. cess rates or improve them than using more distant data. Therefore, As future work we consider including further patient’s details we conclude that the incremental window is a good scheme for the about their admission, such as the antibiotics they have been admin- training set to make predictions of this particular problem. Also in istered, whether they have required intubation, if they have needed these last experiments, in which test years used are 2012 and 2013, mechanical ventilation, among others, which are indicators that can the accuracy values achieved are higher than the ones in previous ex- have an impact on the appearance of resistant bacteria. Also, includ- periments where the accuracy value was accumulated from 2006 to ing past antibiogram information, for each particular patient whose 2013. The fact that accuracy improves as the training set is bigger sensibility test has to be predicted, would be an interesting approach implies that the algorithm is able to learn the model and generalize to evaluate. To extend this study we propose to use other kind of DM from the training data. algorithms different to LR, to assess whether they can improve suc- cess rates seen in this study. In addition, it would be advantageous to Table 7. Year 2012 and year 2013 prediction for 4th Gen. Cephalosporins. predict the resistance to the six antimicrobial families relevant to P. aeruginosa mentioned in the current study, so it would be possible to 2012 2013 detect when a bacteria may become multidrug resistant. Metric 04-11 Inc. 04-11 04-12 Inc. Accuracy 59.57 59.57 75.0 75.0 80.0 Resistant success 15.38 15.38 0.0 0.0 0.0 ACKNOWLEDGEMENTS Sensitive success 76.47 76.47 78.95 78.95 84.21 Average 45.93 45.93 39.47 39.47 42.11 We are thankful to University Hospital of Fuenlabrada, Madrid, In the two left columns, comparison of the year 2012 prediction Spain for providing the database used in this research. using a fixed training window with instances from 2004 to 2011 and This work has been partly supported by the Spanish Thematic an incremental training window. In the three columns on the right, Network “Learning Machines for Singular Problems and Applica- comparison of 2013 prediction using a fixed training window with tions (MAPAS)” (TIN2017-90567-REDT, MINECO/FEDER EU), instances from 2004 to 2011, a fixed training window from 2004 to 2012 and an incremental training window. The target feature predicted by the IDEAI-UPC Consolidated Research Group Grant from Cata- in all cases is c&cf4. lan Agency of University and Research Grants (AGAUR, Generalitat de Catalunya) (2017 SGR 574), by the Spanish Ministry of Economy, Industry and Competitiveness under the Research Project Klinilycs (TEC2016-75361-R), by the Science and Innovation Ministry Grants AAVis-BMR (PID2019-107768RA-I00) and BigTheory (PID2019- 5 CONCLUSIONS AND FUTURE WORK 106623RB-C41), by the Spanish Institute of Health Carlos III (grant In this paper we suggest to use health records and past antibiogram DTS 17/00158), by Project Ref. F656 financed by Rey Juan Carlos data to predict antimicrobial resistance in the ICU. We propose to University, by the Young Researchers R&D Project Ref. 2020-661, use information about recently detected resistant germs in ICU as financed by Rey Juan Carlos University and Community of Madrid features of the data set. To handle changes in data distribution over (Spain), and by the Youth Employment Initiative (YEI) R&D Project time caused by progressive mutation of bacteria, we suggest to use Ref. TIC-11649 financed by the Community of Madrid (Spain). an incremental window for training set, and a test window with fixed size such that training and test instances are temporarily as close as possible. REFERENCES Experiments show that information of recent resistant bacteria de- [1] Gustavo Arango-Argoty, Emily Garner, Amy Pruden, Lenwood S tected in patients of the ICU, contains a relatively high amount of in- Heath, Peter Vikesland, and Liqing Zhang, ‘Deeparg: a deep learning approach for predicting antibiotic resistance genes from metagenomic Dawn M Sievert, ‘Antimicrobial-resistant pathogens associated with data’, Microbiome, 6(1), 1–15, (2018). healthcare-associated infections: summary of data reported to the na- [2] K Balakrishnan, Exponential distribution: theory, methods and appli- tional healthcare safety network at the centers for disease control and cations, Routledge, 2018. prevention, 2011–2014’, infection control & hospital epidemiology, [3] Nele Brusselaers, Dirk Vogelaers, and Stijn Blot, ‘The rising problem of 37(11), 1288–1301, (2016). antimicrobial resistance in the intensive care unit’, Annals of intensive [22] Shichao Zhang, Xindong Wu, and Manlong Zhu, ‘Efficient missing data care, 1(1), 47, (2011). imputation for supervised learning’, in 9th IEEE International Confer- [4] Rafael Cantón, ‘Lectura interpretada del antibiograma: una necesidad ence on Cognitive Informatics (ICCI’10), pp. 672–679. IEEE, (2010). clı́nica’, Enfermedades Infecciosas y microbiologı́a clı́nica, 28(6), 375– 385, (2010). [5] Thomas M Cover and Joy A Thomas. Elements of information theory, 2012. [6] C Defez, P Fabbro-Peray, N Bouziges, A Gouby, A Mahamat, JP Dau- res, and A Sotto, ‘Risk factors for multidrug-resistant pseudomonas aeruginosa nosocomial infection’, Journal of Hospital Infection, 57(3), 209–216, (2004). [7] Milislav Demerec, ‘Origin of bacterial resistance to antibiotics’, Jour- nal of bacteriology, 56(1), 63, (1948). [8] MJ Ellington, O Ekelund, Frank Møller Aarestrup, R Canton, M Doumith, Christian Giske, Hajo Grundman, Henrik Hasman, MTG Holden, Katie L Hopkins, et al., ‘The role of whole genome sequenc- ing in antimicrobial susceptibility testing of bacteria: report from the eucast subcommittee’, Clinical microbiology and infection, 23(1), 2– 22, (2017). [9] Håkan Hanberger, José-Angel Garcia-Rodriguez, Miguel Gobernado, Herman Goossens, Lennart E Nilsson, Marc J Struelens, et al., ‘Antibi- otic susceptibility among aerobic gram-negative bacilli in intensive care units in 5 european countries’, Jama, 281(1), 67–71, (1999). [10] S Joshi et al., ‘Hospital antibiogram: a necessity’, Indian journal of medical microbiology, 28(4), 277, (2010). [11] Lisa L Maragakis, Eli N Perencevich, and Sara E Cosgrove, ‘Clini- cal and economic burden of antimicrobial resistance’, Expert review of anti-infective therapy, 6(5), 751–763, (2008). [12] Sergio Martı́nez-Agüero, Inmaculada Mora-Jiménez, Jon Lérida- Garcı́a, Joaquı́n Álvarez-Rodrı́guez, and Cristina Soguero-Ruiz, ‘Ma- chine learning techniques to identify antimicrobial resistance in the in- tensive care unit’, Entropy, 21(6), 603, (2019). [13] Scott T Micek, Ann E Lloyd, David J Ritchie, Richard M Reich- ley, Victoria J Fraser, and Marin H Kollef, ‘Pseudomonas aeruginosa bloodstream infection: importance of appropriate initial antimicrobial treatment’, Antimicrobial agents and chemotherapy, 49(4), 1306–1311, (2005). [14] Marcus Nguyen, S Wesley Long, Patrick F McDermott, Randall J Olsen, Robert Olson, Rick L Stevens, Gregory H Tyson, Shaohua Zhao, and James J Davis, ‘Using machine learning to predict antimicrobial mics and associated genomic features for nontyphoidal salmonella’, Journal of clinical microbiology, 57(2), (2019). [15] Mitchell W Pesesky, Tahir Hussain, Meghan Wallace, Sanket Patel, Saadia Andleeb, Carey-Ann D Burnham, and Gautam Dantas, ‘Eval- uation of machine learning and rules-based approaches for predicting antimicrobial resistance profiles in gram-negative bacilli from whole genome sequence data’, Frontiers in microbiology, 7, 1887, (2016). [16] Paz Revuelta-Zamorano, Alberto Sánchez, José Luis Rojo-Álvarez, Joaquı́n Álvarez-Rodrı́guez, Javier Ramos-López, and Cristina Soguero-Ruiz, ‘Prediction of healthcare associated infections in an in- tensive care unit using machine learning and big data tools’, in XIV Mediterranean Conference on Medical and Biological Engineering and Computing 2016, pp. 840–845. Springer, (2016). [17] AD Russell, ‘Antibiotic and biocide resistance in bacteria: introduc- tion’, Journal of applied microbiology, 92, 1S–3S, (2002). [18] Jean-Francois Timsit, Stephan Harbarth, and Jean Carlet. De-escalation as a potential way of reducing antibiotic use and antimicrobial resis- tance in icu, 2014. [19] ML Tlachac, Elke A Rundensteiner, Kerri Barton, Scott Troppy, Kirthana Beaulac, and Shira Doron, ‘Predicting future antibiotic susceptibility using regression-based methods on longitudinal mas- sachusetts antibiogram data.’, in HEALTHINF, pp. 103–114, (2018). [20] Alexey Tsymbal, Mykola Pechenizkiy, Padraig Cunningham, and Seppo Puuronen, ‘Handling local concept drift with dynamic integra- tion of classifiers: Domain of antibiotic resistance in nosocomial infec- tions’, in 19th IEEE Symposium on Computer-Based Medical Systems (CBMS’06), pp. 679–684. IEEE, (2006). [21] Lindsey M Weiner, Amy K Webb, Brandi Limbago, Margaret A Dudeck, Jean Patel, Alexander J Kallen, Jonathan R Edwards, and