Knowledge discovery from emergency ambulance dispatch during COVID-19: A case study of Nagoya City, Japan
Essam A. Rashed, Sachiko Kodera, Hidenobu Shirakami, Ryotetsu Kawaguchi, Kazuhiro Watanabe, Akimasa Hirata
KKnowledge discovery from emergency ambulance dispatch duringCOVID-19: A case study of Nagoya City, Japan
Essam A. Rashed a,b , Sachiko Kodera a , Hidenobu Shirakami c , Ryotetsu Kawaguchi c , Kazuhiro Watanabe c ,Akimasa Hirata a,d,e a Department of Electrical and Mechanical Engineering, Nagoya Institute of Technology, Nagoya 466-8555, Japan b Department of Mathematics, Faculty of Science, Suez Canal University, Ismailia 41522, Egypt c Nagoya City Fire Department, Nagoya, Aichi, Japan d Center of Biomedical Physics and Information Technology, Nagoya Institute of Technology, Nagoya 466-8555, Japan e Frontier Research Institute for Information Science, Nagoya Institute of Technology, Nagoya 466-8555, Japan
Abstract
Accurate forecasting of medical service requirements is an important big data problem that is crucial forresource management in critical times such as natural disasters and pandemics. With the global spread ofcoronavirus disease 2019 (COVID-19), several concerns have been raised regarding the ability of medicalsystems to handle sudden changes in the daily routines of healthcare providers. One significant problemis the management of ambulance dispatch and control during a pandemic. To help address this problem,we first analyze ambulance dispatch data records from April 2014 to August 2020 for Nagoya City, Japan.Significant changes were observed in the data during the pandemic, including the state of emergency (SoE)declared across Japan. In this study, we propose a deep learning framework based on recurrent neuralnetworks to estimate the number of emergency ambulance dispatches (EADs) during a SoE. The fusionof data includes environmental factors, the localization data of mobile phone users, and the past historyof EADs, thereby providing a general framework for knowledge discovery and better resource management.The results indicate that the proposed blend of training data can be used efficiently in a real-world estimationof EAD requirements during periods of high uncertainties such as pandemics.
Keywords:
Deep learning, long short-term memory (LSTM), COVID-19, emergency ambulance dispatch(EAD)
1. Introduction
An ambulance is one of the most important health-care tools providing an essential life-saving role on adaily basis. The management and location of theambulance dispatch center are known to reduce therate of fatalities, particularly during a national emer-gency caused by a natural disaster or wide-spread
Email address: [email protected] (Essam A.Rashed) pandemic. The outbreak of the infectious coronavirusdisease 2019 (COVID-19) was reported in China in2019 [1], and the disease spread throughout the globeduring the early months of 2020. Owing to its fastand significant spread, a recognized collapse in medi-cal systems and shortages of medical equipment havebeen reported in several countries [2–4]. Clinicalsymptoms of COVID-19 include fever, cough, anddifficulty breathing, which are not radically differentfrom those of other seasonal infections in their earlystage [5], which increases the challenge for the accu-
Preprint submitted to Elsevier February 18, 2021 a r X i v : . [ c s . A I] F e b ate diagnosis and treatment. To prevent virus infec-tion at hospitals, an early pandemic protocol was con-sidered [6]. Emergency service management shouldbe included in this protocol. Later on, a guidancefor health worker infection prevesion was relased byWorld Health Organization (WHO) [7]. Paramedicshave also suffered from the spread of this novel virus;thus, a more careful management than usual is re-quired [8–10].In Japan, the first state of emergency (SoE) wasdeclared nationwide on April 16, 2020, and was re-voked on May 25, 2020. Note that, during this pan-demic, a complete closure policy was not adopted inJapan, but rather a voluntary segregation and seclu-sion with community cooperation was applied. Dur-ing the SoE, outdoor activities were reduced, particu-larly in common crowded regions such as major trainstations. From April 18, 2020, NTT Docomo, Inc. (amobile phone operator in Japan) started to providepublicly available data on activities based on the userlocations of mobile phones or smartphones in majortraffic stations nationwide.The careful management of healthcare infrastruc-ture is required for fair allocation and usage especiallywhere these resources are limited [11]. Owing to alack of data on ambulance services in such a pandemicera, it was unclear how many allocations were neededfor ambulances, including the number of emergencyambulance dispatches (EADs) and special require-ments for additional processes, such as disinfectionafter the transport of potentially positive cases. Theopen questions here are i) how accurate EAD fore-casts can be when applying environmental factors,ii) how the changes caused by abnormalities such asa SoE during a pandemic should be handled, and iii)what are the main reasons for such changes, includ-ing the relationship with human activities? If suchinformation is made available, additional measuresfor ambulance management can be implemented. Forexample, it will be possible to strictly limit the useof certain ambulance units for potential COVID-19patients.This study first analyzed EAD data recorded fromApril 1, 2014, to August 18, 2020, in Nagoya City,Japan. This analysis has led to a high-quality esti-mation of the number of EADs based on data gath- Figure 1: Map of Japan with Aichi prefecture and Nagoya Cityhighlighted. ered before the COVID-19 era using machine learn-ing approaches. A comparison of the estimated andactual EADs observed during the pandemic clarifiesthe differences caused by the COVID-19 outbreak. Adata analysis model can provide better understand-ing of the potential approaches used to estimate thenumber of EADs during a pandemic and calls duringa SoE. The main factor for this was also discussedfor potential future pandemics, including third wavesof COVID-19. To the best of the authors’ knowl-edge, this is the first study that applies a deep learn-ing approach to forecast the daily number of EADswhen considering environmental factors, even duringnon-pandemic states. The main contributions of thisstudy can be summarized as follows: • A machine learning architecture for accurateEAD forecasting in urban areas from environ-mental factors • A trained long short-term memory (LSTM) net-work for EAD estimation in Nagoya City, Japan,with a potential extension to other regions basedon data availability2
The introduction of a new social factor (i.e., mo-bile phone usage) that can be used to fine-tunethe EAD forecasting during a pandemic
2. Materials and methods
Nagoya is a major city in Japan located in the cen-tral region of Honshu Island (Fig. 1) with a popula-tion varied from 2,272 k to 2,316 k (2014–2020), whichmakes Nagoya the fourth largest city nationwide .All daily EAD data were collected by authorities atthe Fire Department of the City. Fig. 2 illustratesthe daily total number of EADs from April 2014 toAugust 2020. The total number of EADs can bethought of as a U-shaped plot with two annual peaksduring the summer and winter, which is highly asso-ciated with upper and/or lower temperature peaks.Weather data, including maximum and/or minimumdaily temperature and other related factors such ashumidity, were collected for Nagoya City from the on-line resources of the Japan Meteorological Agency .We also processed data representing the variations(in percentage) of people around the major trans-port stations in Japan collected by the mobile ca-reer NTT Docomo, Inc. and released on the web .These data were made available immediately after theemergency declaration on April 18, 2020. Note thatthe national market share of NTT Docomo is approx-imately 36.9% (ranked first) . The data are based onthe estimated statistical population generated frommobile terminal network operational data [12]. It is well known that the number of ambulance dis-patches is related to the daily average ambient tem-perature [13–19]. The effects of the daily maximumand minimum ambient temperatures, as well as therelative humidity, on the number of daily EADs are https://mobaku.jp/covid-19/ (21 Aug. 2020) shown in Fig. 3. A common U-shape curve can beobserved at both the maximum and minimum tem-peratures, whereas the relative humidity triggers aslightly higher number of EADs within the middlerange. However, the effect of environmental factorscan be split into different patterns based on the causeor illness corresponding to the EADs.In addition, weekends and holidays are other fac-tors that characterize human social activities. Fur-thermore, the age of the population affects such ac-tivities. The typical retirement age in Japan is cur-rently approximately 60-65 years. In addition, an ageof over 65 is classified as elderly; thus, 65 years is usedas a reference value in this study. The former factorswere used as input for applying machine learning tothe data on the previous 5 years. The statistics areapplied for the populations younger and older than65 years in age. The learning-based estimation is ap-plied as a reference value for 2020 and then comparedwith the observed value determined by the city firedepartment. Unexpected pandemics are known to cause exten-sive requirements for medical care services and spe-cial resource management [20]. From the data pre-sented in Fig. 4, clearly, the number of EADs in 2020is lower than that during the past 5 years. Figure 5shows the number of EADs according to the maxi-mum daily ambient temperature. Data were dividedinto categories based on the pickup site (i.e., indoor oroutdoor). The normal pattern observed since April2014 significantly changed during the pandemic, assmaller numbers of dispatches were observed in bothlocation categories. A further reduction was also rec-ognized during and after the first SoE (up to August19). One potential reason for the reduction in EADsduring the pandemic is the change in social activi-ties. However, there are several other factors suchas the enhanced hygiene [21] and avoiding the accessto medical facilities to prevent potential COVID-19infection [22]. It is difficult to explicitly highlight themain reason but it is likely a composition of many fac-tors include the above mentioned ones. Even usingmodels that enable artificial intelligence (or machine3 igure 2: Daily total number of recorded EADs in Nagoya City from April 1, 2014 to August 19, 2020. Years are presented indifferent colors.Figure 3: Correlation between the daily total EAD and themaximum (left) and minimum (middle) temperatures and rel-ative humidity (right) using data from April 1, 2014 to Decem-ber 31, 2019. learning), specific data patterns are challenging to es-timate using models trained on data measured undernormal situations.
Recurrent neural networks (RNNs) are a widerange of network architectures that consider inter-neuronal connections such that they formulate amemory-like unit [23]. LSTM is a commonly usedclass of RNN that performs well with long-term de-pendencies (Fig. 6). To address this problem, we con-sider an LSTM-based network architecture. Considera time-series data sample X = ( x (1) , x (2) , . . . , x ( T ) ) (cid:62) ,where x ( t ) = ( x ( t )1 , x ( t )2 , . . . , x ( t ) I ) and x ( t ) i is the mea-surement of the i th variable at the t th time slot frame.The computation within a single LSTM node can beexpressed as follows: i ( t ) = σ ( W ix x ( t ) + W is s ( t − + b i ) , (1) o ( t ) = σ ( W ox x ( t ) + W os s ( t − + b o ) , (2) f ( t ) = σ ( W fx x ( t ) + W fs s ( t − + b f ) , (3) m ( t ) = φ ( W mx x ( t ) + W hc s ( t − + b m ) , (4) c ( t ) = f t (cid:12) c ( t − + i t (cid:12) m ( t − , (5) s ( t ) = o ( t ) (cid:12) φ ( c ( t ) ) , (6)where i ( t ) , o ( t ) , f ( t ) , m ( t ) , and s ( t ) represent the in-put gate, output gate, forget gate, memory gate, andnode state values at time frame t (here, day index),respectively. W and b are the node weights (param-eters) and bias matrices. σ and φ are the sigmoid and tanh functions, respectively. LSTM layers are trainedusing time-series data to optimize the parameters forfuture forecasting tasks.The proposed architecture consists of two LSTMlayers (with 50 and 30 nodes) and a three fully con-nected (FC) layers (with 300, 100 and K nodes). Theinput data vector contains the daily based maximumambient temperature, average relative humidity, anda binary label that identifies working days from non-working days (weekends or national holidays). Thenetwork output is a K nodes representing the numberfuture days to be predicted. The network parameters( W ) and bias ( b ) are initialized with zero matrices.In addition, the Adam algorithm [24] is used for datafitting during training along with the cross-entropyloss function and automatically estimated learningrate. The network architecture was implemented us-ing Wolfram Mathematica ® ver. 12.1 on a worksta-tion with four Intel ® Xeon CPUs running at 3.60GHz, with 128 GB of memory and three NIVIDIAGeForce 1080 GPUs.4 igure 4: Box plot of daily EAD calls during 2020 (January 1 to April 17, April 18 to May 25, and May 26 to August 19)compared with the same time periods from 2015 to 2019. Top row is the percentage of daily AED per population and bottomrow is the AED numbers for elderly (age greater than 65 years) citizens. These time periods are before, during, and after thefirst SoE declaration in Japan to mitigate COVID-19 pandemic. Right-side graph demonstrate the Nagoya city population.Figure 5: Ambulance dispatches in Nagoya City are associatedwith the daily maximum ambient temperature for patients se-lected based on their indoor and outdoor activities. (a) Plotsof all data measured during a normal state (April 2014 to De-cember 2019). (b) and (c) plots of both normal (gray) andafter outbreak for indoor and outdoor calls, respectively. Dataand their regression curve with label (1) refer to (a). Regres-sion curves with labels (2), (3), and (4) indicate the start ofthe COVID-19 pandemic spread but prior to the first SoE dec-laration (January 1, 2020, to April 17, 2020), during the firstSoE (April 18 to 25 May) and after the first SoE (May 26 toAugust 19), respectively. The vertical scale is magnified.
3. Results
We investigated two scenarios that consider thepre- and ongoing pandemic periods. The pre-pandemic results are used to validate the feasibilityof the LSTM architecture in estimating the numberof daily EADs when considering the different cate-gories and conditions. We further analyzed the effectof the Covid-19 pandemic and demonstrated a fine-tuning used to effectively handle bias factors result-ing from such abnormalities. The forecast accuracy isvalidated using the correlation coefficient (CC) meanabsolute error (MAE), which is defined as follows:CC( u, v ) = n (cid:80) i u i v i − (cid:80) i u i (cid:80) i v i (cid:112) [ n (cid:80) i u i − ( (cid:80) i u i ) ][ n (cid:80) i v i − ( (cid:80) i v i ) ] , (7)MAE( u, v ) = 1 n (cid:88) i | u i − v i | u i . (8)Here, u and v are n -size vectors representing thereal and estimated numbers of EADs, respectively.5 igure 6: Left is the network architecture and right is the basic structure of a single LSTM node. We consider data from April 1, 2014, to Decem-ber 31, 2018, for training and 2019 data for testing.The output is split into different categories consid-ering the pickup location (indoor/outdoor) and agecategories: children (0 ≤ age ≤ < age ≤ > Under this scenario, we consider the data fromApril 1, 2014, to December 31, 2019, as the train-ing set and the data from January 1 to August 19,2020, for testing. The network is trained similarly tothe first scenario setup; however, a significant error is found as the number of EADs is significantly de-creased in 2020 (Fig. 8). To overcome this problem,the data on the mobile users’ location are includedas an additional input term. The available data pro-vided by NTT around Nagoya’s main station (April18 to August 18, 2020) are shown in Fig. 9 along withprofile of COVID-19 positive cases in Japan . We as-sume that the mobile phone usage is 100% duringJanuary 2020 where other missing data were linearlyinterpolated. Figure 8 shows the forecast results withand without mobile data for different data groups.We observed a significant improvement in the esti-mation accuracy. Table 2 presents the descriptivestatistics and assessment metrics. These data clearlydemonstrate that mobile data usage as a surrogate ofsocial activities is useful in improving the forecastingof EADs in different data groups. To clearly demon-strate the forecast accuracy, a plot representing theestimated number of daily EADs in association withthe daily maximum temperature and average rela-tive humidity is shown in Fig. 10. Estimated dataare demonstrated with almost the same pattern inboth cases; however, including the mobile users’ lo-cation data significantly reduces the error caused bythe abnormalities. igure 7: Actual and estimated number of EADs during 2019 for (a) all groups and various age groups of (b) children, (c)adults, and (d) the elderly, as well as the location groups of (e) outdoor and (f) indoor patients. able 1: Descriptive statistics and error metrics for different data groups estimated for year 2019.Group Mean Std. Error Median Mode StDev Kurtosis Skewness Range Min Max Sum CC MAEAll Real
Est.
Real
Est.
Real
Est.
Real
Est.
Real
Est.
Real
Est.
Est.+ ) and without (
Est. ) mobiledata usage for year 2020.Group Mean Std. Error Median Mode StDev Kurtosis Skewness Range Min Max Sum CC MAEAll
Real
Est.
Est.+
Real
Est.
Est.+
Real
Est.
Est.+
Real
Est.
Est.+
Real
Est.
Est.+
Real
Est.
Est.+ igure 8: Actual and estimated numbers of EADs in 2020 (January 1 to August 19) for (a) all groups and various age groupsof (b) children, (c) adults, and (d) the elderly, as well as location groups of (e) outdoor and (f) indoor patients. Estimationsthat include mobile usage data provide more accurate results. igure 9: Reduction rate of mobile users around Nagoya main station in year 2020 compared with the values measured in theprevious year. A reduction in usage shows a consistent regression lines with the data split into working and non-working days.A reference curve in gray color demonstrate the number of daily positive COVID-19 cases during first and second waves inJapan (line represents a 7-days average).Figure 10: Actual and estimated number of daily EADs (January 1 to August 19, 2020) associated with maximum temperature(left) and average absolute humidity (right). Curves represent the best-fitting polynomial (3rd degree). Both weather factors,including mobile usage data, improve the forecast accuracy. .3. Contribution of different variables To validate how different variables contribute tothe estimation of daily EADs, we conducted an abla-tion study. The experimental data presented in Fig. 8is repeated with exclusion of single variable each. Weconsider exclusion of mobile usage data, maximumtemperature, average humidity and day label (work-ing day/off day). The network is retrained with thesedifferent set of variables and data of 2020 is estimatedfor each case. Results are shown in Fig. 11 alongwith box plots demonstrate MAE in each case. Theseresults demonstrate that training using all variablesleads to MAE of 4.70% and mobility data that rep-resent a surrogate of social activities during the pan-demic is the most dominant variable that increasethe MAE to 19.12% when excluded. Excluding tem-perature, humidity and day label are of comparableimportance and lead to increase MAE to values of8.44%, 7.66% and 8.04%, respectively.
In some cases, it is required to have a long-termforecasting that demonstrate data beyond just a sin-gle day. The network architecture is designed to ex-press estimation of K successive days. The trainingsession is repeated with different architectures with K = 3 , ,
14 and 28 with all other parameters fixedas those shown in Fig. 8 and the number of EADsfor 2020 is estimated once more. Figure 12 demon-strate results obtained from different value of K . Thefuture K days are computed for each day from thebeginning of 2020. Therefore, each day is presentedwith different values (1 to K ). We demonstrate theaverage, minimum and maximum estimated daily val-ues. It is clear from these data, that a good estima-tion can be achieved within small time period (e.g.3 days), however, estimation error accumulates whenthe estimation period extended further. This is clearfrom the regions labeled with the dashed line ellipsein Fig. 12. MAE associated with K = 3 , ,
14 and 28is 7.32%, 7.79%, 8.02% and 8.46%, respectively.
4. Discussion
In this study, the numbers of people transportedby ambulance and based on population activities were evaluated for the planning of the third waveof COIVD-19 and future pandemics. As one notablefeature in Japan, the government did not lock downthe city, but requested people to apply voluntary con-straints. Nagoya is a primal city in the third largestarea, following the Kanto (Tokyo) and Kansai (Os-aka) areas, in Japan. In addition, ambulance useis free in Japan; thus, the number of EADs corre-sponds approximately to the real number of patientswho need such transport.The number of patients transported by ambulanceduring the state of the emergency was generallysmaller than that during previous years. This differ-ence may be attributable to the significant reductionin the activities of adults; the percentage of telework-ing in Japan is normally approximately 1%, whereasit was 30% in April 2020. As a discrepancy in the ac-tivities of the population, teleworking became com-mon around the city center but not in the suburbs.Patients younger than 65 years in age should bewell correlated with the activities of the populationin the city center. Similarly, a reduction was observedeven in the elderly. We then demonstrated that envi-ronmental factors such as the maximum temperatureand relative humidity can be used to estimate thenumber of people transported by ambulance. Dur-ing a pandemic, special care (e.g., disinfection) is re-quired even for emergency services. During this par-ticular pandemic, ambulances were disinfected whentransporting potential COVID-19 patients, at leastin Nagoya. Thus, resource management was criticalduring the pandemic, and maintaining the number ofdispatches below a certain level has been essential.During the pandemic, the number of patients trans-ported was reduced by 20% at maximum, whereasthe amount of human activity around the central sta-tion was suppressed by 80%. The findings here willbe useful to estimate or plan ambulance allocations.The second SoE in Japan was declared active fromJanuary 7 up to March 7, 2021 (tentative schedule).The average daily mobility reduction rate in the firstand second SoEs and the normal period in betweenat Nagoya main station was 63%, 29.1% and 18.8%, data computed from January 8 to February 14, 2021 igure 11: Left is the actual and estimated number of daily EADs using (a) all data and when exclude (b) mobile usage, (c)temperature, (d) humidity and (e) day label data. Right is the box plots demonstrate MAE for cases of (a)-(e). respectively. This indicates the positive response ofpublic (in different scales) during emergency calls tovoluntarily reduce the social activities.As a limitation of this study, the EAD data werenot classified into specific diseases, some of whichmay be highly related to environmental factors (suchas heat stroke or respiratory system failure), whereasothers may not be related at the same level. However,splitting the data and applying a further generaliza-tion of the proposed model to handle this problemmay remain as a future study. Moreover, a compar-ison with data obtained from different cities wouldprovide a better understanding of the outlined frame-work.The source code used in this study includingtrained network will be provided publicly after pub-lication.
5. Conclusion
This study investigated the correlation between en-vironmental factors such as ambient temperature, ab-solute humidity, and the daily number of EADs inNagoya City, Japan. Data collected from April 2014indicate a good correlation that may be potentiallyuseful in future forecasting of required AED facilitiesbased on weather data. A machine learning frame-work based on an LSTM network architecture wasused for time-sampled forecasting, and interesting re-sults were shown within a normal state. This findingpresents for the first time an affordable method for estimating the number of EADs with environmen-tal factors using the LSTM architecture. Moreover,a strong bias was recognized when forecasting thenumber of EADs required during the COVID-19 pan-demic. To handle this problem, additional data in-dicating a reduction in mobile phone usage in majorcrowded areas, such as train stations, were used as asurrogate for the reduction of social activity duringthe pandemic. Including these data can significantlyreduce the forecasting error during a time of uncer-tainty, such as during unexpected pandemics.
References [1] Z. Wu, J. M. McGoogan, Characteristics of andimportant lessons from the coronavirus disease2019 (COVID-19) outbreak in china: Summaryof a report of 72314 cases from the chinese cen-ter for disease control and prevention, JAMA323 (13) (2020) 1239–1242. doi:10.1001/jama.2020.2648 .[2] B. Armocida, B. Formenti, S. Ussai, F. Palestra,E. Missoni, The italian health system andthe COVID-19 challenge, The Lancet Pub.Health 5 (5) (2020) e253. doi:10.1016/S2468-2667(20)30074-8 .[3] M. L. Ranney, V. Griffeth, A. K. Jha, Criticalsupply shortages - the need for ventilators and12 igure 12: Actual and estimated number of daily EADs in 2020 with forecasting of 3, 7, 14 and 28 days in (a)-(d), respectively.Average, maximum and minimum estimated values are used to demonstrate error range. Highlighted region with dashed ellipseclearly demonstrate that estimation error increase with long forecasting time period. doi:10.1056/NEJMp2006141 .[4] A. Raoofi, A. Takian, A. A. Sari, A. Olyaee-manesh, H. Haghighi, M. Aarabi, COVID-19pandemic and comparative health policy learn-ing in Iran, Arch. Iran. Med. 23 (4) (2020) 220–234. doi:10.34172/aim.2020.02 .[5] C. Sohrabi, Z. Alsafi, N. O’Neill, M. Khan,A. Kerwan, A. Al-Jabir, C. Iosifidis, R. Agha,World health organization declares global emer-gency: A review of the 2019 novel coronavirus(COVID-19), Int. J. Surg. 76 (2020) 71–76. doi:10.1016/j.ijsu.2020.02.034 .[6] W. Glauser, Proposed protocol to keep COVID-19 out of hospitals, CMAJ 192 (10) (2020) E264–E265. doi:10.1503/cmaj.1095852 .[7] World Health Organization, Prevention, iden-tification and management of health workerinfection in the context of COVID-19, 30October 2020, Tech. rep., No. WHO/2019-nCoV/HW infection/2020.1 (2020).[8] S. Whitfield, A. Macquarrie, M. Boyle, Respond-ing to a cardiac arrest: Keeping paramedicssafe during the COVID-19 pandemic, Aust. J.Paramed. 17. doi:10.33151/ajp.17.809 .[9] J. E. Buick, S. Cheskes, M. Feldman, P. R. Ver-beek, M. Hillier, Y. C. Leong, I. R. Drennan,COVID-19: What paramedics need to know!,CJEM (2020) 1–5 doi:10.1017/cem.2020.367 .[10] R. Higginson, A. Parry, M. Williams, B. Jones,Paramedics and pneumonia associated withCOVID-19, J. Paramed. Pract. 12 (5) (2020)179–185. doi:10.12968/jpar.2020.12.5.179 .[11] E. J. Emanuel, G. Persad, R. Upshur, B. Thome,M. Parker, A. Glickman, C. Zhang, C. Boyle,M. Smith, J. P. Phillips, Fair allocation of scarcemedical resources in the time of Covid-19, N.Engl. J. Med. 382 (2020) 2049–2055. doi:10.1056/NEJMsb2005114 . [12] M. Terada, T. Nagata, M. Kobayashi, Popu-lation estimation technology for mobile spatialstatistics, NTT DoCoMo Techn. J. 14 (2013) 10–15.[13] E. Alessandrini, S. Z. Sajani, F. Scotto,R. Miglio, S. Marchesi, P. Lauriola, Emergencyambulance dispatches and apparent tempera-ture: A time series analysis in Emilia-Romagna,Italy, Environ. Res. 111 (8) (2011) 1192–1200. doi:10.1016/j.envres.2011.07.005 .[14] K. L. Bassil, D. C. Cole, R. Moineddin, W. Lou,A. M. Craig, B. Schwartz, E. Rea, The relation-ship between temperature and ambulance re-sponse calls for heat-related illness in Toronto,Ontario, 2005, J. Epidemiol. Commun. H.65 (9) (2011) 829–831. doi:10.1136/jech.2009.101485 .[15] J. Cheng, Z. Xu, D. Zhao, M. Xie, H. Yang,L. Wen, K. Li, H. Su, Impacts of tempera-ture change on ambulance dispatches and sea-sonal effect modification, Int. J. Biometeo-rol. 60 (12) (2016) 1863–1871. doi:10.1007/s00484-016-1173-4 .[16] K. Kotani, K. Ueda, X. Seposo, S. Yasukochi,H. Matsumoto, M. Ono, A. Honda, H. Takano,Effects of high ambient temperature on am-bulance dispatches in different age groups inFukuoka, Japan, Global Health Action 11 (1)(2018) 1437882. doi:10.1080/16549716.2018.1437882 .[17] K. Sangkharat, M. A. Mahmood, J. E. Thornes,P. A. Fisher, F. D. Pope, Impact of extremetemperatures on ambulance dispatches in Lon-don, UK, Environ. Res. 182 (2020) 109100. doi:10.1016/j.envres.2019.109100 .[18] D. Patel, L. Jian, J. Xiao, J. Jansz, G. Yun,T. Lin, A. Robertson, Joint effects of heatwavesand air quality on ambulance services for vul-nerable populations in Perth, western Australia,Environ. Pollut. 252 (2019) 532–542. doi:10.1016/j.envpol.2019.05.125 .1419] J. Hu, Y. Wen, Y. Duan, S. Yan, Y. Liao,H. Pan, J. Zhu, P. Yin, J. Cheng, H. Jiang,The impact of extreme heat and heat waves onemergency ambulance dispatches due to externalcause in Shenzhen, China, Environ. Pollut. 261(2020) 114156. doi:10.1016/j.envpol.2020.114156 .[20] I. Bielajs, F. M. Burkle Jr., F. L. Archer,E. Smith, Development of prehospi-tal, population-based triage-managementprotocols for pandemics, Prehosp Dis-aster Med. 23 (5) (2008) 420–430. doi:10.1017/s1049023x00006154 .[21] E. Cerulli Irelli, B. Orlando, E. Cocchi,A. Morano, F. Fattapposta, V. Di Piero,D. Toni, M. R. Ciardi, A. T. Giallonardo,G. Fabbrini, A. Berardelli, C. Di Bonaven-tura, The potential impact of enhanced hygienicmeasures during the COVID-19 outbreak onhospital-acquired infections: A pragmatic studyin neurological units, Journal of the Neurologi-cal Sciences 418 (2020) 117111. doi:10.1016/j.jns.2020.117111 .[22] S. J. Lange, M. D. Ritchey, A. B. Goodman,T. Dias, E. Twentyman, J. Fuld, L. A. Schieve,G. Imperatore, S. R. Benoit, A. Kite-Powell,et al., Potential indirect effects of the COVID-19 pandemic on use of emergency departmentsfor acute life-threatening conditions—UnitedStates, January–May 2020 (2020). doi:10.15585/mmwr.mm6925e2 .[23] A. Graves, M. Liwicki, S. Fern´andez, R. Berto-lami, H. Bunke, J. Schmidhuber, A novel con-nectionist system for unconstrained handwritingrecognition, IEEE T. Pattern Anal. 31 (5) (2009)855–868. doi:10.1109/TPAMI.2008.137doi:10.1109/TPAMI.2008.137