Hospital emergency demand and congestion estimation method and system based on big data

Abstract

The present invention provides a hospital emergency demand and congestion estimation method based on multi-source urban big data, comprising the following steps: obtaining a taxi drop-off data set, and clustering the drop-off points using a self-organizing mapping network; extracting the drop-off time series characteristics, the first type of spatial characteristics, and the second type of spatial characteristics according to each cluster obtained by clustering, and using collaborative training and active learning strategies to obtain the time series of emergency demand; establishing a piecewise stable queuing model, and dividing the emergency department into general outpatient type, internal medicine outpatient type, and surgical outpatient type; using the queuing model to calculate the average number of queues and the average waiting time under each outpatient type and the occupancy rate of the infusion room, that is, the utilization rate of the infusion bed ρ infusion ; The three types of indicators are integrated to obtain an estimate of the congestion level of the entire emergency department; The method provided by the present invention has the advantages of high efficiency and low consumption, and also achieves a high accuracy rate.

Description

Technical Field

[0001] The present invention relates to the fields of data mining, big data and urban computing, and in particular to a method and system for estimating hospital emergency demand and congestion based on multi-source urban big data. Background Art

[0002] The emergency department of a hospital is an important institution that provides emergency medical services to patients. However, in real life, since no appointment is required in advance, patients often have to wait in long queues, which leads to a poor medical experience. This situation also increases the risk of clustered infections, especially during public health emergencies. Therefore, accurately modeling emergency medical needs and making reasonable estimates of the degree of congestion can not only improve patients' medical experience, but also play an extremely important role in the rational allocation of medical resources.

[0003] Traditional medical needs monitoring methods are mostly divided into several categories. They rely on on-site observations and reports by medical staff, which increases time and labor costs; they rely on video analysis by surveillance cameras, which saves manpower but easily infringes on patient privacy; and they rely on registration information from the hospital information system, which is more accurate but involves privacy and cannot be made public. Summary of the invention

[0004] The main purpose of the present invention is to overcome the above-mentioned defects in the prior art and propose a method for estimating hospital emergency demand and congestion based on multi-source urban big data. The method of the present invention extracts medical emergency demand based on self-organizing mapping, collaborative training and active learning strategies, and estimates congestion based on queuing theory models. It has the advantages of high efficiency and low consumption, and also achieves a high extraction and estimation accuracy.

[0005] The present invention adopts the following technical solution:

[0006] The method for estimating hospital emergency demand and congestion based on multi-source urban big data includes the following steps:

[0007] Get a taxi drop-off dataset and use a self-organizing map network to cluster drop-off points;

[0008] According to each cluster obtained by clustering, the passenger drop-off time series feature, the first type of spatial feature and the second type of spatial feature are extracted, where the first type of spatial feature is the spatial distance feature, and the second type of spatial feature is the surrounding POI feature;

[0009] Using the obtained passenger drop-off time series features, the first-category spatial features, and the second-category spatial features, the collaborative training and active learning strategies are used to obtain the time series of emergency demand.

[0010] A piecewise stable queuing model was established, and emergency departments were divided into general outpatient clinic types, internal medicine outpatient clinic types, and surgical outpatient clinic types;

[0011] Using the queuing model to calculate the average number of people in line Q under the general practice clinic type 1 and the average waiting time W 1 ; Use the queuing model to calculate the average number of people in line Q under the type of internal medicine clinic 2 and the average waiting time W 2 , and record the third indicator: the occupancy rate of the infusion room, that is, the utilization rate of the infusion bed ρ infusion , using the queuing model to calculate the average number of people in line Q under the surgical outpatient type 3 and the average waiting time W 3 ;

[0012] The three types of indicators were integrated to obtain an estimate of crowding for the entire emergency department.

[0013] Specifically, we obtain a taxi drop-off dataset and use a self-organizing map network to cluster drop-off points, including:

[0014] Use Python's pandas package to aggregate the taxi drop-off dataset;

[0015] Call the self-organizing map network provided by the minisom package and use the taxi dataset for training until the maximum number of iterations is reached;

[0016] After training is completed, multiple cluster center coordinates are obtained;

[0017] Determine the cluster center to which each passenger drop-off record belongs, and after traversal, obtain the cluster to which each record belongs, thereby completing the clustering of the entire passenger drop-off data set.

[0018] Specifically, the obtained passenger drop-off time series features, the first type of spatial features, and the second type of spatial features are used to obtain the time series of emergency needs using collaborative training and active learning strategies, including:

[0019] First, determine a labeled dataset G 1 , and the unlabeled dataset is denoted as G 2 ;

[0020] Using the labeled dataset G 1 and the passenger drop-off time series feature F t Train the temporal classifier L t , using the labeled dataset G 1 and spatial feature F s Train the spatial classifier L s ; The spatial features include first-type spatial features and second-type spatial features;

[0021] Using the time classifier L t and the spatial classifier L sFor the unlabeled dataset G 2 The samples in the set are classified, and the samples with the same classification results are added to the first set C 1 , inconsistent addition to the second set C 2 ;

[0022] From the first set C 1 Select high confidence samples with set rankings and add them to the labeled dataset G 1 ;

[0023] From the second set C 2 The samples with low confidence ranking are selected and judged by experts, labeled and added to the labeled dataset G. 1 ;

[0024] Repeat the iterative training process until the unlabeled dataset G 2 is empty, and the clusters related to emergency are obtained;

[0025] Extract the passenger drop-off time series of each emergency-related cluster;

[0026] The passenger drop-off time series of emergency-related clusters within a set range around the hospital are aggregated to obtain the time series of the hospital's emergency visit demand.

[0027] Specifically, the three types of indicators are integrated to obtain the crowding estimate of the entire emergency department, which is:

[0028]

[0029]

[0030] ρ=ρ infusion

[0031] Among them, Q is the number of people queuing in the emergency department, W is the waiting time in the emergency department, and ρ is the utilization rate of emergency infusion beds.

[0032] Another aspect of the present invention provides a hospital emergency demand and congestion estimation system based on multi-source city big data, including:

[0033] Drop-off point clustering unit: obtains taxi drop-off data set and uses self-organizing mapping network to cluster drop-off points;

[0034] Feature extraction unit: extract the passenger drop-off time series feature, the first type of spatial feature and the second type of spatial feature according to each cluster obtained by clustering, where the first type of spatial feature is the spatial distance feature and the second type of spatial feature is the surrounding POI feature;

[0035] Emergency demand time series acquisition unit: using the obtained passenger drop-off time series features, the first-class spatial features and the second-class spatial features, the collaborative training and active learning strategies are used to obtain the time series of emergency demand;

[0036] Queuing model building unit: building a piecewise stable queuing model, and dividing the emergency department into general outpatient type, internal medicine outpatient type and surgical outpatient type;

[0037] Index calculation unit: Use the queuing model to calculate the average number of people in line Q for general outpatient clinics 1 and the average waiting time W 1 ; Use the queuing model to calculate the average number of people in line Q under the type of internal medicine clinic 2 and the average waiting time W 2 , and record the third indicator: the occupancy rate of the infusion room, that is, the utilization rate of the infusion bed ρ infusion , using the queuing model to calculate the average number of people in line Q under the surgical outpatient type 3 and the average waiting time W 3 ;

[0038] Crowding estimation unit: The three types of indicators are integrated to obtain the crowding estimate of the entire emergency department.

[0039] Specifically, in the drop-off point clustering unit, a taxi drop-off data set is obtained, and a self-organizing mapping network is used to cluster the drop-off points, specifically including:

[0040] Use Python's pandas package to aggregate the taxi drop-off dataset;

[0041] Call the self-organizing map network provided by the minisom package and use the taxi dataset for training until the maximum number of iterations is reached;

[0042] After training is completed, multiple cluster center coordinates are obtained;

[0043] Determine the cluster center to which each passenger drop-off record belongs, and after traversal, obtain the cluster to which each record belongs, thereby completing the clustering of the entire passenger drop-off data set.

[0044] Specifically, the obtained passenger drop-off time series features, the first type of spatial features, and the second type of spatial features are used to obtain the time series of emergency needs using collaborative training and active learning strategies, including:

[0045] First, determine a labeled dataset G 1 , and the unlabeled dataset is denoted as G 2 ;

[0046] Using the labeled dataset G 1 and the passenger drop-off time series feature Ft Train the temporal classifier L t , using the labeled dataset G 1 and spatial feature F s Train the spatial classifier L s ; The spatial features include first-type spatial features and second-type spatial features;

[0047] Using the time classifier L t and the spatial classifier L s For the unlabeled dataset G 2 The samples in the set are classified, and the samples with the same classification results are added to the first set C 1 , inconsistent addition to the second set C 2 ;

[0048] From the first set C 1 Select high confidence samples with set rankings and add them to the labeled dataset G 1 ;

[0049] From the second set C 2 The samples with low confidence ranking are selected and judged by experts, labeled and added to the labeled dataset G. 1 ;

[0050] Repeat the iterative training process until the unlabeled dataset G 2 is empty, and the clusters related to emergency are obtained;

[0051] Extract the passenger drop-off time series of each emergency-related cluster;

[0052] The passenger drop-off time series of emergency-related clusters within a set range around the hospital are aggregated to obtain the time series of the hospital's emergency visit demand.

[0053] Specifically, the three types of indicators are integrated to obtain the crowding estimate of the entire emergency department, which is:

[0054]

[0055]

[0056] ρ=ρ infusion

[0057] Among them, Q is the number of people queuing in the emergency department, W is the waiting time in the emergency department, and ρ is the utilization rate of emergency infusion beds.

[0058] On the other hand, an embodiment of the present invention provides a non-transitory computer-readable storage medium storing computer instructions, characterized in that the computer instructions are used to execute the steps of the above-mentioned hospital emergency demand and congestion estimation method based on multi-source urban big data.

[0059] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:

[0060] (1) The present invention provides a hospital emergency demand and congestion estimation method based on multi-source urban big data, obtains a taxi drop-off data set, and uses a self-organizing mapping network to cluster the drop-off points; according to each cluster obtained by clustering, extracts the drop-off time series characteristics, the first type of spatial characteristics, and the second type of spatial characteristics, wherein the first type of spatial characteristics are spatial distance characteristics, and the second type of spatial characteristics are surrounding POI characteristics; using the obtained drop-off time series characteristics, the first type of spatial characteristics, and the second type of spatial characteristics, a collaborative training and active learning strategy are used to obtain the time series of emergency demand; a piecewise stable queuing model is established, and the emergency department is divided into general outpatient type, internal medicine outpatient type, and surgical outpatient type; the queuing model is used to calculate the average number of people in line Q under the general outpatient type 1 and the average waiting time W 1 ; Use the queuing model to calculate the average number of people in line Q under the type of internal medicine clinic 2 and the average waiting time W 2 , and record the third indicator: the occupancy rate of the infusion room, that is, the utilization rate of the infusion bed ρ infusion , using the queuing model to calculate the average number of people in line Q under the surgical outpatient type 3 and the average waiting time W 3 ; The three types of indicators are integrated to obtain the congestion estimate of the entire emergency department; The method of the present invention extracts medical emergency needs based on self-organizing mapping, collaborative training and active learning strategies, and estimates congestion based on queuing theory models. It has the advantages of high efficiency and low consumption, and also achieves a high extraction and estimation accuracy.

[0061] (2) The present invention provides a method for estimating hospital emergency demand and congestion based on multi-source urban big data. The emergency department is divided into general outpatient clinics, internal medicine outpatient clinics and surgical outpatient clinics based on the divide-and-conquer method, and the congestion level is estimated, which is in line with the actual situation and the method is universal. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 An overall flow chart of the technical solution provided by an embodiment of the present invention.

[0063] Figure 2 The SOM cluster center and hospital distribution preview diagram provided by the embodiment of the present invention;

[0064] Figure 3 A diagram of the collaborative training and active learning process provided by an embodiment of the present invention;

[0065] Figure 4The effect diagram of emergency demand extraction based on cluster classification provided by the embodiment of the present invention; wherein (a) is the distribution map of the drop-off points of the First Affiliated Hospital of Xiamen University; (b) the visit time series of emergency patients; (c) the arrival time series of hospital staff;

[0066] Figure 5 A comparison chart of the emergency needs extracted by Xiamen University Affiliated Zhongshan Hospital and the actual medical records of the hospital provided in an embodiment of the present invention;

[0067] Figure 6 A schematic diagram of a general emergency department consultation process provided by an embodiment of the present invention.

[0068] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0069] like Figure 1 This is the overall process framework diagram of the technical solution of the embodiment of the present invention. The technical solution is mainly divided into two-stage frameworks. One is the extraction of the visit demand of the emergency department, and the other is the estimation of the emergency congestion level. In the emergency demand extraction stage, a SOM neural network model based on competitive learning is adopted. After clustering to obtain clusters, collaborative training and active learning are combined with time and space characteristics to classify and extract medical emergency needs. In the emergency congestion estimation stage, based on the "divide-and-conquer" idea, three types of queuing models are integrated to estimate the length of the emergency queue, the average waiting time and the injection room occupancy rate;

[0070] The specific implementation steps are as follows:

[0071] The method for estimating hospital emergency demand and congestion based on multi-source urban big data includes the following steps:

[0072] S1: Get the taxi drop-off dataset and use the self-organizing map network to cluster the drop-off points;

[0073] The python pandas package is used to aggregate the existing three-month taxi drop-off data set in Xiamen (September 2016 and September-October 2017).

[0074] Call the SOM network provided by the minisom package, specify 30x25 neurons, and use the taxi dataset for training until the maximum number of iterations is reached.

[0075] After training, 750 cluster center coordinates are obtained.

[0076] The competition strategy of SOM is used to determine the cluster center of each passenger drop-off record point. After traversal, the cluster to which each record belongs is obtained, thereby completing the clustering of the entire passenger drop-off data set.

[0077] At this point, we have completed the clustering of the taxi dataset and can prepare data for the next classification work. Figure 2 shown.

[0078] S2: Extract the passenger drop-off time series features, the first-class spatial features, and the second-class spatial features from each cluster obtained by clustering. The first-class spatial features are spatial distance features, and the second-class spatial features are surrounding POI features.

[0079] Extract time features: For each cluster, obtain its passenger drop-off time series with a granularity of 1 hour, and then calculate the average passenger drop-off sequence on weekdays (Monday to Friday) and holidays (weekends), denoted as f w (c i ) and f n (c i ), and concatenate to obtain the passenger drop-off time series features [f w (c i ),f n (c i )].

[0080] Extract the first type of spatial features: calculate the road network distance between each cluster and the nearest hospital as the spatial distance feature.

[0081] Extract the second type of spatial features: Extract the number of medical-related POI categories (hospitals, pharmacies) in each cluster and use them as the surrounding POI features.

[0082] S3: Using the obtained passenger drop-off time series features, the first-category spatial features, and the second-category spatial features, collaborative training and active learning strategies are used to obtain the time series of emergency needs;

[0083] After completing the extraction of spatiotemporal features of the clusters, we need to determine whether the clusters are related to hospital emergency visits. The traditional idea is to build a classifier to achieve the distinction. However, the heterogeneity of temporal and spatial features means that we cannot simply treat them the same, but should treat them differently. Moreover, it is not so easy to judge whether the clusters are actually related to the emergency, and we need to rely on the experience of experts or internal staff. In order to solve the difficulties of feature heterogeneity and lack of labels, this technical solution uses collaborative training and active learning methods to integrate the previously extracted spatiotemporal features and minimize expert experience to achieve cluster classification and ultimately complete emergency demand extraction.

[0084] (1) Collaborative training

[0085] Co-training is a multi-view learning method that assumes that each data can be classified from different views. Different classifiers can be trained from different angles. These classifiers trained from different angles are then used to classify unlabeled samples. Unlabeled samples that are considered credible are then selected and added to the training set.

[0086] In this technical solution, we use the time feature (F t ) and spatial characteristics (F s ) Train two classifiers L t and L s Specifically, we first use crowdsending to determine a small-scale labeled dataset G 1 , and a large number of unlabeled data sets are recorded as G 2 Then use G 1 Train two classifiers separately:

[0087] L t ←L t· learn(F t ,G 1 )L s ,←L s· learn(F s ,G 1 )

[0088] Next, we will L t and L s Applied to each group G 2 , and select N with the same label and the highest confidence predicted by the two classifiers c samples, and add them to the labeled dataset G 1 We repeat the iterative training process until G 2 Is empty or the maximum number of iterations has been reached.

[0089] (2) Active learning strategy

[0090] The purpose of active learning is to obtain a better performing classifier by manually annotating fewer training samples. It uses machine learning algorithms to query the most useful unlabeled samples, which are labeled by experts, and then these samples are used to retrain the classifier to better improve accuracy.

[0091] In the collaborative training process of this technical solution, two classifiers are applied to each group G 2 A batch of sample sets with inconsistent classification results will be generated. We select a small number of samples with the largest differences and submit them to experts (experienced taxi drivers or traffic police, etc.) for judgment. The returned samples are added to the labeled dataset G 1, and repeat the iterative training process until G 2 Is empty or the maximum number of iterations has been reached.

[0092] like Figure 3 As shown, it is a schematic diagram of the above algorithm process.

[0093] After completing the classification of the above clusters’ relevance to emergency, we extracted the emergency demand of the hospital: first, we extracted the time series of the drop-off of each emergency-related cluster, and then aggregated the time series of the drop-off of emergency-related clusters within a certain range around each hospital to obtain the time series of the hospital’s emergency visit demand. Figure 4 The effect diagram of emergency demand extraction based on cluster classification provided by the embodiment of the present invention; wherein (a) is the distribution map of the drop-off points of the First Affiliated Hospital of Xiamen University; (b) the visit time series of emergency patients; (c) the arrival time series of hospital staff.

[0094] like Figure 5 As shown, a comparison chart of the emergency needs extracted by Xiamen University Affiliated Zhongshan Hospital and the actual medical records of the hospital provided by an embodiment of the present invention.

[0095] S4: Establish a piecewise stable queuing model and divide the emergency departments into general outpatient type, internal medicine outpatient type and surgical outpatient type;

[0096] In the stage of estimating the congestion level of the emergency department, we need to fully understand and model the service process of the emergency department based on the actual situation, and then estimate the congestion level. Since the emergency department in real life is a complex system, we cannot model it with a single queuing model. However, its service types can be roughly divided into three categories: general outpatient clinics generally follow the principle of first-come-first-served, internal medicine outpatient clinics generally have an additional infusion process, and surgical outpatient clinics will give priority to patients with serious conditions, which may include a preemptive process. Therefore, we can adopt the idea of ​​"divide and conquer". First, we build three queuing models for these three common service processes and estimate their parameters. Then, we combine the three queuing models according to the actual situation to estimate the overall congestion level index. We will integrate and provide three indicators: the number of patients waiting in the emergency department, the expected waiting time, and the occupancy rate of the infusion room.

[0097] There are three basic parameters in the queue model, namely, the arrival rate, the service rate, and the number of servers. For the emergency department, all three parameters are time-varying because the patient flow varies over time, the number of doctors varies according to the hospital shift schedule, and the service time varies for patients with different symptoms. Therefore, we decided to use a non-stationary M t / M t / s t The model is used to simulate the queuing process, where the arrival process follows a Poisson distribution, with the arrival rate λ(t), the service process follows an exponential distribution, with the service rate μ(t) and the number of working doctors at time t is s(t).

[0098] Since the non-stationary queuing model requires a lot of data and calculations to estimate parameters, it is not feasible. By observing the emergency demand and service-related parameters contained in the hospital public data set extracted previously, we found that the piecewise stable M / M / s model can be used to approximate the non-stationary model, and real data is used for correction in each time period to simplify the calculation. In addition, for the convenience of modeling, we put forward the following assumptions:

[0099] i) Once a patient enters the queue, they will not leave the queue;

[0100] ii) the waiting room is unlimited, i.e. no patient will be denied entry into the queue;

[0101] iii) the doctor must complete the current treatment before leaving under the shift arrangement;

[0102] At this point, we can use a stationary model to approximate the non-stationary queue model, where the arrival rate is distributed with Poisson mean λ and the service rate is exponentially distributed with mean μ.

[0103] When modeling, we divide the entire time period into P intervals [t i , t i+1 ), i = 1, ..., P, and then the queue in each interval is considered as a stationary model. A notable feature of the emergency department is the alternating periods of underload and overload, that is, sometimes there are long queues waiting, and sometimes some doctors are idle. These underloads and overloads have an impact on subsequent intervals and should be taken into account when modeling. Therefore, consecutive intervals are not completely independent, and we should modify the arrival rate of each time interval based on the previous interval.

[0104] For i th Interval [t i , t i+1 ), let λ i is the actual arrival rate, μ i is the service rate, si is the number of doctors in service. i Defined as th The rate of patients not receiving services during the time interval, while Z i is the sum of the arrival rate in this time interval and the rate of patients not served in the previous time interval, then Z i+1 =D i +λ i+1 , where D i This can be approximated using the Erlang B formula, which calculates the probability that a client cannot be served because all servers are busy. We will Defined as the server utilization rate, P i (B) is defined as the patient th The probability of not receiving treatment at the end of the interval, then we have D i =Z i ·P i (B), where P i (B) can be calculated as follows:

[0105]

[0106] Finally, let λ i modified As i th The actual arrival rates of departments after interval adjustment are:

[0107]

[0108] We will use this result in subsequent crowding estimation.

[0109] In this section, we introduce how to use the hospital emergency demand extracted in the previous section and the acquired hospital public data to calculate the inputs required for the queuing model: arrival rate, service time, and number of doctors.

[0110] Here we clarify the representation and meaning of several parameters:

[0111] Arrival rate λ t :This is the i calculated in the previous part th Interval [t i , t i+1 )’s hospital emergency needs.

[0112] Service rate μ t :This is exponentially distributed in i th Interval [t i , t i+1 ) service rates. We calculated this based on each hospital's standard operating schedule.

[0113] Number of doctors t :This is in i th Interval [t i , t i+1 ), based on open hospital shift schedules.

[0114] In light of the actual situation, the emergency department usually does not require an appointment in advance. The consultation process includes arrival, waiting, diagnosis in the corresponding department, dispatch to the infusion room, referral or joint consultation, and departure. Figure 6 As shown. This technical solution focuses on analyzing situations where people gather and there is a high risk of infection. Therefore, we do not consider referral and joint consultation procedures, and only examine the crowding in public areas, namely waiting rooms and infusion rooms. In order to achieve universality, we consider three common types of outpatient clinics: general outpatient clinics, internal medicine outpatient clinics, and surgical outpatient clinics. For convenience, λ is used to represent the arrival rate λ in any time interval. modified , the service rate is μ, the number of doctors in service is s, and we only need to analyze the queuing process within one interval. Next, we will explain the modeling of each type of crowding indicator.

[0115] S5: Use the queuing model to calculate the average number of people queuing for general outpatient clinics Q 1 and the average waiting time W 1 ; Use the queuing model to calculate the average number of people in line Q under the type of internal medicine clinic 2 and the average waiting time W 2 , and record the third indicator: the occupancy rate of the infusion room, that is, the utilization rate of the infusion bed ρ infusion , using the queuing model to calculate the average number of people in line Q under the surgical outpatient type 3 and the average waiting time W 3 ;

[0116] (1) General Practice Outpatient Clinic:

[0117] The characteristic of general outpatient clinics is that the service time is relatively long, but the patients leave immediately after diagnosis. The specific rule of this model is that all patients are served on a first-come, first-served basis. The service process of such clinics includes 1) patients arriving, 2) waiting in line when all doctors are busy, 3) waiting in line for treatment, and 4) leaving the clinic. We apply the modified M / M / s model to calculate the required indicators. ρ is And ρ<1. All doctors are busy P b The probability is:

[0118]

[0119] The average number of people in line Q and the average waiting time W are:

[0120]

[0121] (2) Internal Medicine Clinic:

[0122] The characteristic of the internal medicine clinic is that the patient may need to stay in the hospital for infusion after the consultation, and may go back to consult the doctor again after the infusion. We assume that the consultation process has s 1 Doctors, service rate μ 1 , the infusion process has s 2 beds, infusion time is where s 2 than 1 The procedure of such an outpatient clinic can be described as follows: 1) patients arrive at the internal medicine clinic at a rate of λ, 2) patients leave each room at the same rate, 3) all patients must first receive a diagnosis, and 4) the probability p of a patient leaving the emergency room directly after the first consultation 1 , 5) Patients have p after the first consultation 2 The probability of staying in the infusion room, 6) patients who are hospitalized for infusion have p 3 The probability of leaving the emergency room, 7) patients who stay in hospital for infusion have p 4 8) The patient must leave or transfer to another department after the second consultation. i , i = 1, 2, 3, 4, which can be found from the public data of the hospital. The total arrival rate of the clinic is λ 1 , the total arrival rate of the infusion room is λ 2 , we have the following equality relationship:

[0123] 1) p 1 +p 2 =p 3 +p 4 =1

[0124] 2)λ 1 =λ+λ 2 p 4

[0125] 3)λ 2 =λp 2

[0126] 4)λ=λ 1 p 1 +λ 2 p 3 ;

[0127] Therefore, we can convert M / M / s 1 The model is applied to the diagnostic process and M / M / s 2 The model is applied to the infusion process. The probability that all doctors are busy is P bi , the waiting time is W i and the queue length Q iThe expression of is the same as that of general practice clinic, where i=1 indicates the diagnosis process, and i=2 indicates the infusion process.

[0128] Next, we can get the queue length Q and waiting time W of the internal medicine clinic:

[0129] Q=Q 1 +Q 2 ,W=W 1 +W 2

[0130] In addition to this, we also calculated and recorded the occupancy rate of the infusion room, that is, the utilization rate of infusion beds, as the third crowding indicator, as follows.

[0131]

[0132] (3) Surgical Clinic

[0133] The characteristic of surgical outpatient clinics is that they give priority to patients with serious conditions. They may not follow the first-come-first-served principle, and there may be a preemptive situation. Therefore, we divide patients into two categories. The first category (urgent patients) takes precedence over the second category (non-urgent patients), that is, patients with low priority can be interrupted by patients with high priority. The service rate of each group is μ 1 , μ 2 , the arrival rate is λ 1 , 2 , which can be obtained from existing data. Since the patients in the first group have priority, their queuing process is not affected by preemption and follows the M / M / s process, where The calculation of Q and W is the same as above.

[0134] The situation of the second group of patients is slightly more complicated because they may be preempted. We define the state (i, j) as the outpatient clinic has i first-class patients and j second-class patients, which are in an idle state, denoted as (i+j), where i+j<s. If all doctors are busy, the second-class patients in this state will either complete the treatment or be interrupted by the first-class patients. Let P be the probability that the second-class patients are not preempted before being served, p ij is the probability that the second type of patient arrives at state (i, j). Using the total probability formula, the probability that the second type of patient needs to wait is:

[0135]

[0136] Using the same M / M / s model and Q and W formulas as before, we can get the queue length and waiting time of the second type of patients, and then integrate them to get the two crowding indicators of the clinic. So far, we have completed the separate analysis of the three types of clinics and the construction of the adaptive queuing model.

[0137] S6: The three types of indicators are integrated to obtain an estimate of the crowding of the entire emergency department.

[0138] Finally, we can integrate these results to estimate the overall crowding and occupancy of the entire emergency department. In real life, each emergency department clinic has a different service process, and patients only go to the clinic that meets their medical needs, that is, the queue processes of different clinics are separated and independent of each other. Let λ be the total arrival rate, and the patient has probability p i Visit th type of outpatient clinic, i = 1, 2, 3, we can use the existing data to obtain this probability. th The outpatient visit rate is λ·p i , each type of clinic can be calculated separately. From the perspective of the entire emergency department, the total number of waiting patients (Q overall ) is the sum of the queue lengths of all outpatient clinics, and the average waiting time of all outpatient clinics (W overall ) can represent the average waiting time of patients. Considering that the infusion room is the most crowded place in the emergency department besides the waiting room, we also recorded the occupancy rate of the infusion room (ρ infusion ), the calculation formula can be referred to the internal medicine outpatient section, and will not be repeated here. In summary, the other two emergency department overall crowding indicators are:

[0139]

[0140]

[0141] ρ=ρ infusion

[0142] Among them, Q is the number of people queuing in the emergency department, W is the waiting time in the emergency department, and ρ is the utilization rate of emergency infusion beds.

[0143] For hospital managers, these three crowding indicators can be used to observe the hospital's operating load to assist in resource allocation and management. For patients who are making decisions, these indicators can provide a reference for their hospital selection.

[0144] Another aspect of the present invention provides a hospital emergency demand and congestion estimation system based on multi-source city big data, including:

[0145] Drop-off point clustering unit: obtains taxi drop-off data set and uses self-organizing mapping network to cluster drop-off points;

[0146] Feature extraction unit: extract the passenger drop-off time series feature, the first type of spatial feature and the second type of spatial feature according to each cluster obtained by clustering, where the first type of spatial feature is the spatial distance feature and the second type of spatial feature is the surrounding POI feature;

[0147] Emergency demand time series acquisition unit: using the obtained passenger drop-off time series features, the first-class spatial features and the second-class spatial features, the collaborative training and active learning strategies are used to obtain the time series of emergency demand;

[0148] Queuing model building unit: building a piecewise stable queuing model, and dividing the emergency department into general outpatient type, internal medicine outpatient type and surgical outpatient type;

[0149] Index calculation unit: Use the queuing model to calculate the average number of people in line Q for general outpatient clinics 1 and the average waiting time W 1 ; Use the queuing model to calculate the average number of people in line Q under the type of internal medicine clinic 2 and the average waiting time W 2 , and record the third indicator: the occupancy rate of the infusion room, that is, the utilization rate of the infusion bed ρ infusion , using the queuing model to calculate the average number of people in line Q under the surgical outpatient type 3 and the average waiting time W 3 ;

[0150] Crowding estimation unit: The three types of indicators are integrated to obtain the crowding estimate of the entire emergency department.

[0151] Specifically, in the drop-off point clustering unit, a taxi drop-off data set is obtained, and a self-organizing mapping network is used to cluster the drop-off points, specifically including:

[0152] Use Python's pandas package to aggregate the taxi drop-off dataset;

[0153] Call the self-organizing map network provided by the minisom package and use the taxi dataset for training until the maximum number of iterations is reached;

[0154] After training is completed, multiple cluster center coordinates are obtained;

[0155] Determine the cluster center to which each passenger drop-off record belongs, and after traversal, obtain the cluster to which each record belongs, thereby completing the clustering of the entire passenger drop-off data set.

[0156] Specifically, the obtained passenger drop-off time series features, the first type of spatial features, and the second type of spatial features are used to obtain the time series of emergency needs using collaborative training and active learning strategies, including:

[0157] First, determine a labeled dataset G 1 , and the unlabeled dataset is denoted as G 2 ;

[0158] Using the labeled dataset G1 and the passenger drop-off time series feature F t Train the temporal classifier L t , using the labeled dataset G 1 and spatial feature F s Train the spatial classifier L s ; The spatial features include first-type spatial features and second-type spatial features;

[0159] Using the time classifier L t and the spatial classifier L s For the unlabeled dataset G 2 The samples in the set are classified, and the samples with the same classification results are added to the first set C 1 , inconsistent addition to the second set C 2 ;

[0160] From the first set C 1 Select high confidence samples with set rankings and add them to the labeled dataset G 1 ;

[0161] From the second set C 2 The samples with low confidence ranking are selected and judged by experts, labeled and added to the labeled dataset G. 1 ;

[0162] Repeat the iterative training process until the unlabeled dataset G 2 is empty, and the clusters related to emergency are obtained;

[0163] Extract the passenger drop-off time series of each emergency-related cluster;

[0164] The passenger drop-off time series of emergency-related clusters within a set range around the hospital are aggregated to obtain the time series of the hospital's emergency visit demand.

[0165] Specifically, the three types of indicators are integrated to obtain the crowding estimate of the entire emergency department, which is:

[0166]

[0167]

[0168] ρ=ρ infusion

[0169] Among them, Q is the number of people queuing in the emergency department, W is the waiting time in the emergency department, and ρ is the utilization rate of emergency infusion beds.

[0170] On the other hand, an embodiment of the present invention provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to execute the above-mentioned hospital emergency demand and congestion estimation method based on multi-source urban big data.

[0171] Computer instructions (also referred to as programs, software, software applications or code) include machine instructions for a programmable processor and may be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as machine-readable signals.

[0172] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A hospital emergency demand and congestion estimation method based on multi-source urban big data, characterized by: The steps include: Get a taxi drop-off dataset and use a self-organizing map network to cluster drop-off points; According to each cluster obtained by clustering, the passenger drop-off time series feature, the first type of spatial feature and the second type of spatial feature are extracted, where the first type of spatial feature is the spatial distance feature, and the second type of spatial feature is the surrounding POI feature; Using the obtained passenger drop-off time series features, the first-category spatial features, and the second-category spatial features, the collaborative training and active learning strategies are used to obtain the time series of emergency demand. A piecewise stable queuing model was established, and emergency departments were divided into general outpatient clinic types, internal medicine outpatient clinic types, and surgical outpatient clinic types; The queuing model is used to calculate the average number of people in line Q1 and the average waiting time W1 for general outpatient clinics; the queuing model is used to calculate the average number of people in line Q2 and the average waiting time W2 for internal medicine outpatient clinics, and the third indicator is recorded: the occupancy rate of the infusion room, that is, the utilization rate of the infusion bed ρ infusion , use the queuing model to calculate the average number of people in line Q3 and the average waiting time W3 under the surgical clinic type; The three types of indicators were integrated to obtain an estimate of crowding for the entire emergency department; Get a taxi drop-off dataset and use a self-organizing map network to cluster drop-off points, including: Use Python's pandas package to aggregate the taxi drop-off dataset; Call the self-organizing map network provided by the minisom package and use the taxi dataset for training until the maximum number of iterations is reached; After training is completed, multiple cluster center coordinates are obtained; Determine the cluster center of each drop-off record point, and obtain the cluster to which each record belongs after traversal, so as to complete the clustering of the entire drop-off data set; Using the obtained passenger drop-off time series features, the first-category spatial features, and the second-category spatial features, the collaborative training and active learning strategies are used to obtain the time series of emergency needs, including: First, determine a labeled dataset G1 and record the unlabeled dataset as G2; Using the labeled dataset G1 and the passenger drop-off time series feature F t Train the temporal classifier L t , using the labeled dataset G1 and spatial features F s Train the spatial classifier L s ; The spatial features include first-type spatial features and second-type spatial features; Using the time classifier L t and the spatial classifier L s Classify the samples in the unlabeled dataset G2, add the samples with consistent classification results to the first set C1, and add the samples with inconsistent classification results to the second set C2; Select high-confidence samples with set rankings from the first set C1 and add them to the labeled dataset G1; Select the low-confidence samples from the second set C2, have them judged by experts, and add them to the labeled dataset G1 after being labeled; Repeat the iterative training process until the unlabeled data set G2 is empty, and obtain the clusters related to emergency; Extract the passenger drop-off time series of each emergency-related cluster; Aggregate the drop-off time series of emergency-related clusters within the set range around the hospital to obtain the time series of emergency visit demand in the hospital; The three types of indicators were integrated to obtain an estimate of the crowding of the entire emergency department, specifically: p=p infusion Among them, Q is the number of people queuing in the emergency department, W is the waiting time in the emergency department, and ρ is the utilization rate of emergency infusion beds.

2. Hospital emergency demand and congestion estimation system based on multi-source urban big data, characterized by: include: Drop-off point clustering unit: obtains taxi drop-off data set and uses self-organizing mapping network to cluster drop-off points; Feature extraction unit: extract the passenger drop-off time series feature, the first type of spatial feature and the second type of spatial feature according to each cluster obtained by clustering, where the first type of spatial feature is the spatial distance feature and the second type of spatial feature is the surrounding POI feature; Emergency demand time series acquisition unit: using the obtained passenger drop-off time series features, the first-class spatial features and the second-class spatial features, the collaborative training and active learning strategies are used to obtain the time series of emergency demand; Queuing model building unit: building a piecewise stable queuing model, and dividing the emergency department into general outpatient type, internal medicine outpatient type and surgical outpatient type; Indicator calculation unit: Use the queuing model to calculate the average number of people in line Q1 and the average waiting time W1 for general outpatient clinics; use the queuing model to calculate the average number of people in line Q2 and the average waiting time W2 for internal medicine outpatient clinics, and record the third indicator: the occupancy rate of the infusion room, that is, the utilization rate of the infusion bed ρ infusion , use the queuing model to calculate the average number of people in line Q3 and the average waiting time W3 under the surgical clinic type; Crowding estimation unit: Integrate the three types of indicators to obtain the crowding estimate of the entire emergency department; In the drop-off point clustering unit, a taxi drop-off data set is obtained, and a self-organizing map network is used to cluster the drop-off points, specifically including: Use Python's pandas package to aggregate the taxi drop-off dataset; Call the self-organizing map network provided by the minisom package and use the taxi dataset for training until the maximum number of iterations is reached; After training is completed, multiple cluster center coordinates are obtained; Determine the cluster center of each drop-off record point, and obtain the cluster to which each record belongs after traversal, so as to complete the clustering of the entire drop-off data set; Using the obtained passenger drop-off time series features, the first-category spatial features, and the second-category spatial features, the collaborative training and active learning strategies are used to obtain the time series of emergency needs, including: First, determine a labeled dataset G1 and record the unlabeled dataset as G2; Using the labeled dataset G1 and the passenger drop-off time series feature F t Train the temporal classifier L t , using the labeled dataset G1 and spatial features F s Train the spatial classifier L s ; The spatial features include first-type spatial features and second-type spatial features; Using the time classifier L t and the spatial classifier L s Classify the samples in the unlabeled dataset G2, add the samples with consistent classification results to the first set C1, and add the samples with inconsistent classification results to the second set C2; Select high-confidence samples with set rankings from the first set C1 and add them to the labeled dataset G1; Select the low-confidence samples from the second set C2, have them judged by experts, and add them to the labeled dataset G1 after being labeled; Repeat the iterative training process until the unlabeled data set G2 is empty, and obtain the clusters related to emergency; Extract the passenger drop-off time series of each emergency-related cluster; Aggregate the drop-off time series of emergency-related clusters within the set range around the hospital to obtain the time series of emergency visit demand in the hospital; The three types of indicators were integrated to obtain an estimate of the crowding of the entire emergency department, specifically: p=p infusiom Among them, Q is the number of people queuing in the emergency department, W is the waiting time in the emergency department, and ρ is the utilization rate of emergency infusion beds.

3. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to execute the method of claim 1.

Images