Road crash emergency response management method and system using a deep hybrid attention framework

Abstract

The present invention is in the field of an emergency response method and system for road crash emergencies, in particular wherein a severity of the road crash is established, an emergency thereof is established, and an emergency response is established, such as by sending an ambulance to the road crash location, or refraining therefrom.

Description

[0001] ROAD CRASH EMERGENCY RESPONSE MANAGEMENT METHOD AND SYSTEM

[0002] USING A DEEP HYBRID ATTENTION FRAMEWORK

[0003] FIELD OF THE INVENTION

[0004] The present invention is in the field of an emergency response method and system for road crash emergencies, in particular wherein a severity of the road crash is established, an emergency thereof is established, and an emergency response is established, such as by sending an ambulance to the road crash location, or refraining therefrom.

[0005] BACKGROUND OF THE INVENTION

[0006] Road traffic crash is a global tragedy that leads to economic loss, injury, and fatalities. Understanding the severity of a road crash at the early stages is vital to timely providing emergency medical services to crash victims. Road traffic crashes are one of the top ten major causes of death worldwide. According to World Health Organization (WHO), road traffic crashes are a major cause for fatalities and human injuries, involving huge economic losses yearly. The vision zero policy that proposes eliminating the chance of severe injury and fatality from road crashes has been considered a target to reach by many agencies worldwide, and they have been striving to diminish road traffic crash likelihood and severe consequences, such as severe injuries and fatalities.

[0007] A road crash may lead to various consequences that range from property damage only to severe injury and fatalities. The severe injury crash requires Emergency Medical Services (EMS) to save road crash victims' lives. Unfortunately, many people lose their lives because of delayed EMS to the crash site, as delay in EMS arrival is one of the significant causes of fatality. Timely providing EMS to crash sites can effectively decrease the probability of fatalities and severe injuries when a crash involves injury. The EMS system needs a proper framework for instant action after the crash.

[0008] Some documents relate to crash severity prediction. For instance Mansoor et al. (DOI: 10.1109 / ACCESS.2020.3040165) recite two-layer ensemble machine learning model is proposed in this study to predict road traffic crash severity. The first layer integrates four base machine learning models: k-nearest neighbour, decision tree, adaptive boosting, and support vector machine; the second layer classifies the crash severity based on the feedforward neural network model. The models are developed using road traffic crash data of road intersections over 6 years (2011-2016) obtained from Great Britain's Department of Transport online database. Only the crash features that can be instantaneously and easily obtained are used as an input. Kashifi et al. (D01: 10.1016 / j.ijtst.2022.07.003) uses a state-of- the-art deep learning-based model that incorporates spatiotemporal information for the shortterm crash prediction, named as Deep Spatiotemporal Hybrid Network (DSHN). The model integrates Convolutional Neural Network (CNN), Long Short-term Memory (LSTM), and Artificial Neural Network (ANN) to incorporate the synergistic power of individual models. The study utilizes different data sources such as big traffic data collected from Paris road network sensors, weather conditions, infrastructure, holidays, and crash data. Liu et al. (DOL 10.1016 / J.TRC.2016.05.023) use a double standard model for allocating limited emergency medical service vehicle resources ensuring service reliability, along with a genetic algorithm (GA), for solving the emergency medical service (EMS) vehicle allocation problem that ensures acceptable service reliability with limited vehicle resources. Without loss of generality, the model is formulated to address emergency services to human injuries caused by vehicle crashes at intersections within an urban street network. The present invention relates in particular to an improved method for emergency response in case of traffic accidents and various aspects thereof which overcomes one or more of the above disadvantages, without jeopardizing functionality and advantages.

[0009] SUMMARY OF THE INVENTION

[0010] The present invention relates in a first aspect to a road crash emergency response method, comprising detecting a road crash, establishing an actual time, in particular an actual time of the road crash, and [approximate] location of road crash, and in particular, based on actual time, (al) establishing temporal conditions, more in particular wherein temporal conditions are selected from at least two of (i) Month, (ii) Day of week, (iii) Time in hour, (iv) Holiday and non-holiday, from (v) lighting conditions, in particular wherein lighting condition is selected from (va) Day-light, (vb) Twilight and dawn, (vc) Night without public lighting, (vd) Night with public lighting off, (ve) Night with public lighting on, and (vf) other lighting conditions, from weather condition (vi), in particular wherein weather condition is selected from (via) Normal, (vib) Light rain, (vic) Heavy rain, (vid) Snow and hail, (vie) Fog and smoke, (vif) Strong wind and storm, (vig) Dazzling weather, (vih) Cloudy weather, and (vij) Other weather condition, and (a2) establishing spatial conditions, wherein spatial conditions comprise at least two first road attributes at the location of the road crash, in particular at least nine road attributes, wherein road attributes are selected from road attribute types, wherein road attribute types are selected from (1) Category of road, in particular wherein category of road is selected from (la) Highway, (lb) National Road, (1c) Departmental Road, (Id) Communal Way, (le) Off-Road, (If) Parking area, and (1g) other road category, from (2) Traffic regime, in particular wherein traffic regime is selected from (2a) One way, (2b) Bidirectional, (2c) Separated carriageways, (2d) Variable lanes, (2e) limited maximum speed roads, in particular 30 or 50 km / h, (2f) medium maximum speed roads, in particular 60 or 80 km / h, (2g) roads with separate pedestrian side-walks, (2h) roads with separate bicycle lanes, and (2i) other traffic regime, from (3) Number of road lanes, from (4) Longitudinal road profile, in particular wherein longitudinal road profile is selected from (4a) Flat, (4b) Slope, (4c) Hilltop, (4d) Hill bottom or valley, and (4e) other longitudinal road profile, from (5) Width of roadway, from (6) Type of facility, in particular wherein Type of facility is selected from (6a) Out of intersection (midblock), (6b) Intersection in an X shape, (6c) Intersection in T shape, (6d) Intersection in Y shape, (6e) Intersection with more than four legs, (6f) Spiral intersection, (6g) Interchange, and (6h) Other intersection, from (7a) State, (7b) Province, (7c) Department, and (7d) region, from (8) Localization, in particular wherein localization is selected from (8a) Rural, and (8b) Urban, from (9) Region, in particular wherein region is selected from (9a) Metropole, (9b) island, such as Antilles, (9c) territory, such as Guyana, (9c) Reunion, and (9d) Mayotte, and from (10) others, in particular from (10a) road material, such as tarmac, asphalt, brick, and stone, (b) feeding spatial conditions and temporal conditions to a data model, using artificial intelligence, and road crash data in a database, establishing an emergency, wherein artificial intelligence comprises use of a dynamic deep hybrid network, (c) based on the emergency, triaging a consequence of said road crash therewith establishing emergency of response, based on the emergency of response, dispatching an emergency vehicle to the location of the road crash or refraining therefrom. An EMS response in road crash incidents is provided (see Fig 1). The process initiates with identifying the crash (step 1), followed by retrieving crucial information such as weather conditions and infrastructure attributes. The time and approximate location of the crash enable the collection of weather data and infrastructure details, respectively (step 2). The present method therefore in particular relates to a unique emphasize on both the temporal and spatial conditions of a road crash, which may be processed separately through distinct input nodes in an Al-driven model. The present method involves in Step al : Establishing temporal conditions based on the actual time of the crash, and using this data to determine relevant temporal features such weather conditions. This means we acquire data that can easily be obtained once we establish the crash time. Thus, there will be no delay in receiving data that will cause the delay in emergency response; and in step a2: Establishing approximate spatial conditions, which specifically involve assessing at least two first road attributes (e.g., road type, width, surface conditions), with a preferred embodiment involving at least nine road attributes. An approximate location of the crash is established, and this will be used to obtain the relevant input features that are immediately available once the approximate location of the crash is known. Once the input features are gathered, they are fed into the proposed model (step 3), which accurately classifies the severity of the crash, distinguishing between severe and non-severe injuries. In the event of a severe injury, the EMS or trauma centres are promptly alerted and dispatched to the crash site (step 4). This may be referred to as triaging. The present method therefore includes a specialized triaging step based on the combined output of the temporal and spatial data processed through an artificial intelligence model. The Al model may integrate crash data from a database and assesses the emergency response needed. The triaging step is designed to determine whether emergency dispatch is necessary or not, based on conditions directly related to the temporal and spatial inputs. This comprehensive framework ensures efficient and timely emergency care services for those in need. Inventors provide a dynamic deep hybrid network called the Deep Hybrid Attention Network (DHAN) for crash severity classification (step 3 in Fig 1). This model leverages the strengths of advanced deep learning techniques, including Long Short-term Memory (LSTM), Deep Neural Network (DNN), and attention mechanisms, to accurately predict the severity of crash events using basic information obtained from the crash site. The DHAN is typically a unique architecture where various layers, including attention mechanisms, work together to dynamically prioritize and process critical information within the network. It typically introduces attention mechanisms or layers. These attention layers allow the model to focus selectively on the most relevant spatial and temporal features, dynamically adjusting the weighting of these features based on their significance to the crash scenario. This aspect is important for achieving high performance in real-time emergency response situations, where not all input data carries equal importance. The DHAN is designed to be more efficient and context-sensitive than traditional ensemble methods. By integrating specialized layers within a unified architecture, the model can more accurately predict complex scenarios by considering both spatial and temporal dynamics in tandem. The attention mechanism further refines this process by enhancing the network’s ability to prioritize relevant information, leading to faster and more reliable predictions in critical applications such as emergency response. Recognizing that combining distinct deep learning layers within a single hybrid architecture, as opposed to using multiple independent models in an ensemble, leads to significant improvements in performance, scalability, and applicability in emergency response systems. This shift in design philosophy and implementation represents a novel solution to the limitations of ensemble methods. , The primary input variables required from the crash site are the time and approximate location of the crash, which can be easily obtained through any existing crash detection mechanism. By integrating the approximate crash location with road asset management data, additional location-related features such as road type and number of lanes can be determined. The temporal features are derived from the recorded crash time. One notable advantage of the present method is its data availability. As the crash time and approximate location are typically available in any crash scenario, the model can be readily applied to different areas or regions. This reduces the challenges associated with model transferability, as data collection methods may vary among different crash recording agencies. By combining the power of deep learning models and attention mechanisms, our approach enhances the accuracy and reliability of crash severity classification. The present invention provides valuable insights for efficient emergency response management systems and can significantly contribute to improving road safety. The present inventors developed a crash emergency response management framework that requires basic crash information for emergency response decision-making. A Deep Hybrid Attention Network (DHAN) was proposed that captures temporal variations and spatial correlations for dynamic severity prediction. Further, two alternative model architectures that initially required only the approximate location or time of the crash were developed and compared with the DHAN. The experiment was conducted on seven years French road crash dataset (2011-2017). The DHAN achieved an AUC of 0.820, an accuracy of 0.761, a recall of 0.803, and a false alarm rate of 0.258, outperforming baseline models. The present invention provides a comprehensive framework that enhances road crash EMS response management. Prompt dispatching of EMS personnel in the event of severe injuries ensures timely and appropriate medical intervention. The present framework holds immense potential to improve the overall effectiveness of road crash emergency response systems, leading to enhanced outcomes and the preservation of human lives. A hybrid deep learning model was used to classify severe and non-severe injuries for deciding the need for EMS after a crash. The model exploits the unique predictive power of deep learning models by combining the LSTM layers parallel to the dense layers, followed by an attention mechanism. The method requires only the approximate location and time, as any crash detection method makes these two variables easy to achieve from the crash site. In addition, the temporal dimension of features was expanded to consider complete temporal information and national holiday data. The present deep learning method was evaluated based on the French crash dataset. The result indicated the superiority of DHAN over the state-of-art methods from the literature. Further, a sensitivity analysis was conducted to evaluate the importance of temporal and spatial features. The analysis indicated that both temporal and spatial features significantly contribute to the model prediction power. Therewith the trained model predicts the probability of a severe injury (or binary decision of severe or non-severe injury crash if you convert those probabilities into a decision) in a crash based on the basic information of a road crash. Based on the input features of the database, the trained model makes a binary output that predicts whether the crash is a severe injury or not severe injury. In case a severe injury is involved, the emergency response (i.e., ambulance) is dispatched to the crash site. Further information can be obtained from the attached document. In summary, the present method is designed to operate effectively with just the time and location of the crash, making it possible to assess the dispatch of emergency services immediately, without waiting for additional crash information. This approach ensures that the emergency response is as swift as possible, leveraging the limited but crucial data available at the earliest stage. Furthermore, the technical contribution of separating the temporal and spatial inputs into dedicated processing nodes is not merely a matter of different feature selection, but a novel architecture that leads to enhanced model performance. This distinction is non-obvious given the trend in prior art toward increasing input complexity rather than simplifying it. The present method is scalable, deployable in various settings, and reliable even in scenarios where detailed crash information is unavailable. This directly addresses and solves the critical problem of delays in emergency response.

[0011] In a second aspect the present invention relates to a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the following instructions: detecting a road crash, establishing an actual time, in particular an actual time of the road crash, and location of road crash, and in particular, based on actual time, (al) establishing temporal conditions, and (a2) establishing spatial conditions, (b) feeding spatial conditions and temporal conditions to a data model, using artificial intelligence, and road crash data in a database, establishing an emergency, (c) based on the emergency, triaging a consequence of said road crash therewith establishing emergency of response, based on the emergency of response, dispatching an emergency vehicle to the location of the road crash or refraining therefrom.

[0012] In general, an Al system categorizes data, typically by receiving a plurality of data points corresponding to said sensory output state and input data; determining one or more subsets of data points from the plurality of data points, wherein said one or more subsets of data points are indicative of said sensory output state and input data, categorizing said sensory output state and input data, said computer program disregarding less or irrelevant data. A computing device may identify patterns using the machine learning algorithm to optimize sensory output. Categorization may involve identifying to which of a set of categories a new captured input may belong, on the basis of a set of training data with known categories, such as the aforementioned categories. Categorization of the one or more subsets of data points associated with a captured input may be performed using one or more machine learning algorithms and statistical classification algorithms. Example algorithms may include linear classifiers (e.g. Fisher's linear discriminant, logistic regression, naive Bayes, and perceptron), support vector machines (e.g. least squares support vector machines), clustering algorithms (e.g. k-means clustering), quadratic classifiers, multi-class classifiers, kernel estimation (e.g. k-nearest neighbour), boosting, decision trees (e.g. random forests), neural networks, Gene Expression Programming, Bayesian networks, hidden Markov models, binary classifiers, and learning vector quantization. Other example classification algorithms are also possible. The process of categorization may involve the computing device determining, based on the output of the comparison of the one or more subsets with the one or more predetermined sets of scene types, a probability distribution (e.g. a Gaussian distribution) of possible scene types associated with the one or more subsets. Those skilled in the art will be aware that such a probability distribution may take the form of a discrete probability distribution, continuous probability distribution, and / or mixed continuous-discrete distributions. Other types of probability distributions are possible as well.

[0013] In an exemplary embodiment of the present method artificial intelligence is based on a convolutional neural network or a neural network with a transformer architecture. In an exemplary embodiment of the present method artificial intelligence is trained on labelled data.

[0014] Thereby the present invention provides a solution to one or more of the above-mentioned problems.

[0015] Advantages of the present description are detailed throughout the description. References to the figures are not limiting, and are only intended to guide the person skilled in the art through details of the present invention.

[0016] DETAILED DESCRIPTION OF THE INVENTION

[0017] In an exemplary embodiment of the present road crash emergency response method the artificial intelligence has two input nodes, a first input node for temporal conditions, and a second input node for road attributes, and wherein the artificial intelligence has at least one output node, the at least one output node configured for providing emergency of response.

[0018] In an exemplary embodiment of the present road crash emergency response method artificial intelligence comprises use of a dynamic deep hybrid network, in particular a deep hybrid attention network (DHAN).

[0019] In an exemplary embodiment of the present road crash emergency response method first input node for temporal conditions of the DHAN comprises at least one of strength levelling, a Long Short-Term Memory (LSTM), a first dense layer processor, and Deep Neural Network(DNN) training, configured for capturing temporal long-term dependencies within temporal conditions from sequential data, and / or configured to establish an emergency.

[0020] In an exemplary embodiment of the present road crash emergency response method spatial conditions are processed through a series of at least two second dense layers, each second dense layer within a spatial branch connecting the at least two second dense layers, configured for extracting spatial crash conditions, in particular weighted spatial crash conditions.

[0021] In an exemplary embodiment of the present road crash emergency response method the artificial intelligence removes dropouts, also considered to relate to outliers, in particular in both first and second input nodes. The dropout layer removes the neurons that are fed from the previous layer randomly. The purpose of dropout layer is to avoid overfitting of the model, which means the model may perform well in training but poorly in testing based on unseen data.

[0022] In an exemplary embodiment of the present road crash emergency response method the outputs of the first and second input nodes are combined into a combined node input, in particular are concatenated, and / or wherein the output of the first input node is “flattened”, that is wherein the output is normalized and / or wherein outliers are removed therefrom.

[0023] In an exemplary embodiment of the present road crash emergency response method the node inputs are processed through an attention module, in particular a combined node input, the attention module configured for attributing an importance level to the respective spatial conditions and temporal conditions, in particular weighted importance, more in particular normalized weighted importance.

[0024] In an exemplary embodiment of the present road crash emergency response method the attention module is configured to process a least one of the combined node input through at least one third dense layer, to flatten processed input, to activate processed input, to repeat processing, to permute processed input, and to multiply processed input with concatenated input. In an exemplary embodiment of the present road crash emergency response method configured to process multiplied input through at least one fourth dense layer, and providing sigmoid output, the sigmoid output configure for providing emergency of response. [The sigmoid function purpose is to make a dichotomy decision (severe injury or non severe injury crash) based on the probabilities of severe injury that is given to this function from the previous layer.

[0025] In an exemplary embodiment of the present road crash emergency response method the LSTM comprises LSTM sub-units, wherein LSTM sub-units are interconnected, wherein LSTM sub-units are in connection to at least one data gate for controlling data flow, each individual sub-unit configured to transfer data to a subsequent layer and to exchange data to a neighbouring sub-unit, and configured to support storage and access of data in the input nodes, and / or comprising at least one first gate for discarding data, at least one second gate for regulating data flow to be transferred to a second gate, and at least one third output gate configured to regulate the data transfer from the previous unit to the next unit of LSTM. The subsequent layer may be any layer that comes after the LSTM. In my proposed model the LSTM is followed by Dropout layer, so the subsequent layer is Dropout in this case. The second gate is the input get and this gate regulates information transfer from the input data to the current unit of LSTM.

[0026] In an exemplary embodiment of the present road crash emergency response method the attention module is configured to categorize concatenated spatial conditions and temporal conditions based on weighted importance of said conditions, and, based on, the respective condition weights obtaining an attention weight, and long term dependencies of spatial conditions and temporal conditions, establishing a crash severity, optionally normalizing attention weights and / or multiplying attention weights with output weights of an immediate previous layer of attention.

[0027] In an exemplary embodiment of the present road crash emergency response method dense layers are configured to be interconnected, in particular by at least one node between two dense layers.

[0028] In an exemplary embodiment of the present road crash emergency response method establishing emergency of response comprises establishing casualties, and severity of said casualties.

[0029] In an exemplary embodiment of the present road crash emergency response method the emergency of response is selected from severe crash, non-severe crash, unknown severity of crash, and uncertain severity of crash.

[0030] In an exemplary embodiment of the present road crash emergency response method at least one of (al) establishing temporal conditions, (a2) establishing spatial conditions (b) feeding spatial conditions and temporal conditions to a data model, and (c) triaging, is carried out by a computer program, in particular a computer program comprising artificial intelligence. In an exemplary embodiment of the present road crash emergency response method each spatial condition is provided with a weight, in particular a normalized weight, more in particular a normalized weight per spatial condition type, such as given in table 1, and / or wherein each temporal condition is provided with a weight, in particular a normalized weight, more in particular a normalized weight per temporal condition type, such as given in table 1.

[0031] The invention is further detailed by the accompanying figures and examples, which are exemplary and explanatory of nature and are not limiting the scope of the invention. To the person skilled in the art it may be clear that many variants, being obvious or not, may be conceivable falling within the scope of protection, defined by the present claims.

[0032] SUMMARY OF FIGURES

[0033] Figures 1, and 2a-f, show details.

[0034] DETAILED DESCRIPTION OF FIGURES

[0035] The figures are further detailed in the description.

[0036] Fig 1 : Proposed framework for road traffic crash emergency response management.

[0037] Fig 2: DHAN Architecture.

[0038] Fig 3: LSTM architecture.

[0039] Fig 4: The original dataset severity levels descriptions and the new severity levels for the emergency response system.

[0040] Figure 5: Crash severity index variations over hours for holidays, weekdays, and weekends.

[0041] Fig 6: Crash frequency (a) and SI (b) distribution over months.

[0042] Fig. 7: Alternative architectures with limited data categories.

[0043] The invention although described in detailed explanatory context may be best understood in conjunction with the accompanying figures.

[0044] Experimental results

[0045] A. DHAN Architecture

[0046] The proposed DHAN architecture, depicted in Fig 2, is specifically designed to address the traffic crash severity problem. The DHAN architecture consists of two input nodes and one output node, each independently handling temporal and spatial correlations in the crash severity. To capture temporal dependencies within the crash features, the temporal branch of the DHAN incorporates an LSTM layer. LSTM is a suitable choice for extracting long-term dependencies in the temporal dimension of the features, enabling the model to capture crucial patterns over time. Conversely, the spatial features are processed through a series of dense layers within the spatial branch. These layers facilitate the extraction of relevant spatial information from the crash features. The outputs of the temporal and spatial branches are then concatenated and passed through an attention mechanism. The attention mechanism allows the model to focus on important features while suppressing irrelevant ones, enhancing the overall performance of the DHAN. Following the attention mechanism, the concatenated features are further processed through a dense layer, and the final output is generated using a sigmoid activation function in the output node. Detailed descriptions of each layer in the DHAN model and its learning process are provided in the subsequent sections, outlining the specific mechanisms and steps employed to effectively predict traffic crash severity.

[0047] B. LSTM Unit

[0048] LSTM is a robust deep learning architecture that is specifically designed to effectively capture and process information from sequential data. Its property in processing and retaining information over long sequences makes it highly applicable to the context of crash severity prediction. The LSTM architecture, as illustrated in Fig 3, is composed of interconnected LSTM units that not only transfer information to the subsequent layer but also exchange information among neighbouring units. This inter-unit communication facilitates the flow of information in a chain-like fashion within the temporal dimension of the data, represented as timesteps. In the context of crash severity prediction, the sequential flow of information within each LSTM unit allows the nodes to store and access information from past sequences. This capability enables the LSTM architecture to capture the temporal dependencies inherent in crash data. By leveraging the stored information from past sequences, LSTM units are able to make predictions about the status of crash severity.

[0049] In the context of crash severity prediction, LSTM units play a vital role in regulating the flow of information to neighbouring units and subsequent layers. These units utilize three gates to control the information flow, allowing for effective processing and retention of relevant crash severity information. The first gate that is known as forget gate, regulates the extent of information that should be transferred to the current cell state from the previous cell state. By controlling this transfer, the forget gate helps in managing the retention of relevant crash severity information. The input gate regulates the flow of information that is received from input features at the time (t) to be transferred to the cell state. It determines which information from the input features is essential for capturing crash severity patterns and incorporates it into the cell state. Conversely, the output gate regulates the flow of information from the past hidden state to the current hidden state. By regulating the output gate, the LSTM unit determines how much information from the past hidden state should be passed on to the current hidden state, considering its relevance in predicting crash severity. The operator regulates the amount of information from the input gate to be used to update the cell state. The current hidden state is finally updated by information from the output gate and cell state.

[0050] C. Attention Mechanism

[0051] Although the LSTM and its parallel and series combination with other deep learning layers as hybrid models capture long-term dependencies, they fail to prioritize the input features for learning. The present attention mechanism addresses this problem. The attention mechanism applies feature importance weight distribution on input features of the respective spatial conditions and temporal conditions, to emphasize essential features and to neglect redundant features. It also considers the long-term dependency on temporal features. The longterm temporal dependencies of data are found to play a significant role in crash severity prediction; however, all of the past historical information is not essential for future prediction of crash severity. The attention mechanism is applied after concatenating temporal and spatial branches (Fig 2). If the attention parameters, typically obtained during backpropagation of data, are calculated from the attention weights coefficients. The attention weights are normalized) and then multiplied by the hidden state of the output from the immediate previous layer of attention to achieve the weighted output representation of the attention layer. The attention weights and biases are trained similarly to other parameters in the model. The weights and biases are updated iteratively during backpropagation until the optimum values are achieved.

[0052] D. Fully connected layer, merging layer, and model learning

[0053] A fully connected layer (Dense) is constructed by connecting nodes called neurons to all subsequent layer nodes. Each neuron processes input information by applying a non-linear function and transfers it to the subsequent layer by assigning weights and a bias after activating it by an activation function. The weights and biases are updated during model training. In the specific case of crash severity prediction, the output layer of the network is activated by using the sigmoid activation function. A sigmoid is chosen as output layer activation because that outputs the crash as severe or non-severe injury, and the sigmoid function is appropriate for binary output. To avoid overfitting, the dropout mechanism was applied between layers. The dropout layer randomly removes neurons during model training. After assembling the DHAN by merging spatial and temporal branches, the model is trained by supervised backpropagation. The model continuously updates weights and biases in each layer to minimize the loss in each iteration.

[0054] IV. DATA ACQUISITION AND PREPARATION

[0055] The present inventors used a France's road crash dataset with data between 2011 and 2017, containing 426,262 road crash records (a public dataset available at https: / / www.onisr.securite-routiere.gouv.fr / en / data-tools / open-data). The data for 2011-2015 was allocated for training our model. The rest of the data was separated for validation (2016 data with 59,432 samples) and testing (2017 data with 60,701 samples) models. The datasets were supplemented with additional data, such as weather conditions and road asset management information, to enrich the dataset. The original dataset contained four severity levels: 1. Unscathed crashes with property damage only; 2. Light injuries are those crashes that led to at least one light injury to users without hospitalization; 3. Hospitalized injuries are those crashes that have led to at least one severe injury requiring hospitalization; and 4. Fatal injury crashes involve at least one fatality within one month of the crash occurrence. The unscathed and light injury crashes were combined as non-severe injuries, and hospitalized and fatal crashes as severe injuries (Fig 4).

[0056] Table 1 shows feature description, categories and normalized weights.

[0057] No Feature description Coded Categories Percentage (%)

[0058] Spatial Features

[0059] 1 Category of road catr Highway 8.1

[0060] National Road 8.0

[0061] Departmental Road 34.9

[0062] Communal Way 46.9

[0063] Off-Road 0.2

[0064] Parking lot 0.4 other 1.4

[0065] 2 Traffic regime traffic One way 17.1

[0066] Bidirectional 64.9

[0067] Separated carriageways 13.7

[0068] Variable lanes 4.3

[0069] 3 Number of road lanes nlane NA

[0070] 4 Longitudinal profile prof Flat 76.4

[0071] Slope 14.6

[0072] Hilltop 3.0

[0073] Hill bottom 4.0

[0074] 5 Width of roadway rwidth NA

[0075] 6 Type of facility int Out of intersection (midblock) 69.8

[0076] Intersection in an X shape 12.6

[0077] Intersection in T shape 9.5

[0078] Intersection in Y shape 1.6

[0079] Intersection with more than four legs 1.1 Spiral intersection 0.7

[0080] Interchange 0.1

[0081] Other intersection 1.7

[0082] 7 Department dep 1-105 NA

[0083] 8 Localization agg Rural 64.5

[0084] Urban 35.5

[0085] 9 Region reg Metropole 98.5

[0086] Antilles 0.8

[0087] Guyana 0.5

[0088] Reunion 0.2

[0089] Mayotte 0.1

[0090] Temporal Features

[0091] 1 Month month January - December NA 2 Day of week weekday Saturday - Friday NA

[0092] 3 Time in hour time 1-24 NA

[0093] 4 Holiday holiday Not holiday 97.8

[0094] Holiday 2.2

[0095] 5 Lighting conditions light Day 68.9

[0096] Twilight or dawn 6.1

[0097] Night without public lighting 8.4

[0098] Night with public lighting off 0.8 Night with public lighting on 15.8

[0099] 6 Weather conditions weath Normal 80.7

[0100] Light rain 10.4

[0101] Heavy rain 2.2

[0102] Snow - hail 0.6

[0103] Fog - smoke 0.8

[0104] Strong wind - storm 0.3

[0105] Dazzling weather 1.0

[0106] Cloudy weather 3.4

[0107] Other 0.7

[0108] Weights could be determined using a Gini impurity model.

[0109] Various in depth analysis on the data have been carried out. Fig 5 indicates the temporal variation of the crash severity index for hours of the day, weekdays, weekends, and holidays. Fig 6 indicates Crash frequency (a) and SI (b) distribution over months. Also the average crash severity index spatial heterogeneity over 101 France departments were investigated. The severity index is widely heterogeneous over departments, varying from 0.3 up to 0.9. Hence, the departmental severity variations and the rest of the spatial patterns significantly impact the model prediction performance. The spatial features were analysed by the sequence of dense layers and incorporated separately into DHAN.

[0110] The present DHAN method was compared against five baseline models to evaluate its performance.

[0111] Table 2: Performance comparison

[0112] Method AUC Accuracy Recall FAR

[0113] DHAN 0.8201± 0.001 0.7607± 0.001 0.8032± 0.002 0.2584± 0.0026

[0114] The present DHAN method performed better in terms of area under the curve (AUC), Accuracy, and Recall, and was comparable in terms of False Alarm Rate (FAR).

[0115] For a better interpretation of the DHAN model performance, the prediction accuracy for each department of France was studied. The accuracy for random guesses in binary classification is 0.5, so less than this value indicates prediction failures. None of the departments has failed prediction. Only one department has an accuracy close, while the rest of the departments indicate satisfactory prediction performance. In an alternative model of fig. 7 training the DHAN model was achieved solely on temporal features. This scenario evaluated the model's performance by exclusively considering the temporal branch of DHAN and excluding the spatial branch. To capture long-term temporal variations, LSTM layers were applied to the temporal data. Subsequently, an attention mechanism was employed to prioritize the model's training based on long-term temporal dependencies. The DHAN-spatial (Fig 7), aimed to retain the spatial branch of DHAN while training the model solely on spatial features. In DHAN-spatial, two dense layers were applied to the spatial features. These were followed by an attention mechanism, a dense layer, and a sigmoid activation as the final layer. Table 3 summarizes the results.

[0116] Table 3: DHAN performance with different feature sets

[0117] Scenario AUC Accuracy Recall FAR

[0118] DHAN 0.8201±0.001 0.7607±0.001 0.8032±0.002 0.2584±0.0026

[0119] DHAN-temp. 0.7930±0.001 0.7511±0.001 0.7854±0.003 0.2610±0.0024

[0120] DHAN-spatial 0.7967±0.001 0.7396±0.001 0.7391±0.002 0.2682±0.0021

[0121] It should be appreciated that for commercial application it may be preferable to use one or more variations of the present system, which would be similar to the ones disclosed in the present application and are within the spirit of the invention.

Claims

CLAIMS1. A road crash emergency response method, comprising detecting a road crash, establishing an actual time, in particular an actual time of the road crash, and location of road crash, and in particular, based on actual time, (al) establishing temporal conditions, wherein temporal conditions are selected from at least two of (i) Month, (ii) Day of week, (iii) Time in hour, (iv) Holiday and non-holiday, from (v) lighting conditions, in particular wherein lighting condition is selected from (va) Day-light, (vb) Twilight and dawn, (vc) Night without public lighting, (vd) Night with public lighting off, (ve) Night with public lighting on, and (vf) other lighting conditions, from weather condition (vi), in particular wherein weather condition is selected from (via) Normal, (vib) Light rain, (vic) Heavy rain, (vid) Snow and hail, (vie) Fog and smoke, (vif) Strong wind and storm, (vig) Dazzling weather, (vih) Cloudy weather, and (vij) Other weather condition, and(a2) establishing spatial conditions, wherein spatial conditions comprise at least two first road attributes at the location of the road crash, in particular at least nine road attributes, wherein road attributes are selected from road attribute types, wherein road attribute types are selected from (1) Category of road, in particular wherein category of road is selected from (la) Highway, (lb) National Road, (1c) Departmental Road, (Id) Communal Way, (le) Off-Road, (If) Parking area, and (1g) other road category, from (2) Traffic regime, in particular wherein traffic regime is selected from (2a) One way, (2b) Bidirectional, (2c) Separated carriageways, (2d) Variable lanes, (2e) limited maximum speed roads, in particular 30 or 50 km / h, (2f) medium maximum speed roads, in particular 60 or 80 km / h, (2g) roads with separate pedestrian side-walks, (2h) roads with separate bicycle lanes, and (2i) other traffic regime, from (3) Number of road lanes, from (4) Longitudinal road profile, in particular wherein longitudinal road profile is selected from (4a) Flat, (4b) Slope, (4c) Hilltop, (4d) Hill bottom or valley, and (4e) other longitudinal road profile, from (5) Width of roadway, from (6) Type of facility, in particular wherein Type of facility is selected from (6a) Out of intersection (midblock), (6b) Intersection in an X shape, (6c) Intersection in T shape, (6d) Intersection in Y shape, (6e) Intersection with more than four legs, (6f) Spiral intersection, (6g) Interchange, and (6h) Other intersection, from (7a) State, (7b) Province, (7c) Department, and (7d) region, from (8) Localization, in particular wherein localization is selected from (8a) Rural, and (8b) Urban, from (9) Region, in particular wherein region is selected from (9a) Metropole, (9b) island, such as Antilles, (9c) territory, such as Guyana, (9c) Reunion, and (9d) Mayotte, and from (10) others, in particular from (10a) road material, such as tarmac, asphalt, brick, and stone,(b) feeding spatial conditions and temporal conditions to a data model, using artificial intelligence, and road crash data in a database, establishing an emergency, wherein artificial intelligence comprises use of a dynamic deep hybrid network,(c) based on the emergency, triaging a consequence of said road crash therewith establishing emergency of response, based on the emergency of response, dispatching an emergency vehicle to the location of the road crash or refraining therefrom.

2. The road crash emergency response method according to claim 1, wherein the artificial intelligence has two input nodes, a first input node for temporal conditions, and a second input node for road attributes, and wherein the artificial intelligence has at least one output node, the at least one output node configured for providing emergency of response.

3. The road crash emergency response method according to claim 1 or 2, wherein artificial intelligence comprises use of a deep hybrid attention network (DHAN).

4. The road crash emergency response method according to claim 3, wherein first input node for temporal conditions of the DHAN comprises at least one of strength levelling, a Long Short-Term Memory (LSTM), a first dense layer processor, and Deep Neural Net- work(DNN) training, configured for capturing temporal long-term dependencies within temporal conditions from sequential data, and / or configured to establish an emergency.

5. The road crash emergency response method according to any of claims 1-4, wherein spatial conditions are processed through a series of at least two second dense layers, each second dense layer within a spatial branch connecting the at least two second dense layers, configured for extracting spatial crash conditions, in particular weighted spatial crash conditions.

6. The road crash emergency response method according to any of claims 1-5, wherein the artificial intelligence removes dropouts, in particular in both first and second input nodes.

7. The road crash emergency response method according to any of claims 1-6, wherein the outputs of the first and second input nodes are combined into a combined node input, in particular are concatenated, and / or wherein the output of the first input node is flattened.

8. The road crash emergency response method according to claim 7, wherein the node inputs are processed through an attention module, in particular a combined node input, the attention module configured for attributing an importance level to the respective spatial conditions and temporal conditions, in particular weighted importance, more in particular normalized weighted importance.

9. The road crash emergency response method according to claim 8, wherein the attention module is configured to process a least one of the combined node input through at least one third dense layer, to flatten processed input, to activate processed input, to repeat processing, to permute processed input, and to multiply processed input with concatenated input.

10. The road crash emergency response method according to any of claims 2-9, configured to process multiplied input through at least one fourth dense layer, and providing sigmoid output, the sigmoid output configure for providing emergency of response.

11. The road crash emergency response method according to any of claims 4-10, wherein theLSTM comprises LSTM sub-units, wherein LSTM sub-units are interconnected, wherein LSTM sub-units are in connection to at least one data gate for controlling data flow, each individual sub-unit configured to transfer data to a subsequent layer and to exchange data to a neighbouring sub-unit, and configured to support storage and access of data in the input nodes, and / or comprising at least one first gate for discarding data, at least one second gate for regulating data flow to be transferred to a second gate, and at least one third output gate configured to regulate the data transfer from the previous unit to the next unit of LSTM.

12. The road crash emergency response method according to any of claims 8-11, wherein the attention module is configured to categorize concatenated spatial conditions and temporal conditions based on weighted importance of said conditions, and, based on, the respective condition weights obtaining an attention weight, and long term dependencies of spatial conditions and temporal conditions, establishing a crash severity, optionally normalizing attention weights and / or multiplying attention weights with output weights of an immediate previous layer of attention.

13. The road crash emergency response method according to any of claims 8-12, wherein dense layers are configured to be interconnected, in particular by at least one node between two dense layers.

14. The road crash emergency response method according to any of claims 1-13, wherein establishing emergency of response comprises establishing casualties, and severity of said casualties.

15. The road crash emergency response method according to any of claims 1-14, wherein the emergency of response is selected from severe crash, non-severe crash, unknown severity of crash, and uncertain severity of crash.

16. The road crash emergency response method according to any of claims 1-15, wherein at least one of (al) establishing temporal conditions, (a2) establishing spatial conditions (b) feeding spatial conditions and temporal conditions to a data model, and (c) triaging, is carried out by a computer program, in particular a computer program comprising artificial intelligence.

17. The road crash emergency response method according to any of claims 1-16, wherein each spatial condition is provided with a weight, in particular a normalized weight, more in particular a normalized weight per spatial condition type, and / or wherein each temporal condition is provided with a weight, in particular a normalized weight, more in particular a normalized weight per temporal condition type.

18. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the following instructions: detecting a road crash, establishing an actual time, in particular an actual time of the road crash, and location of road crash, and in particular, based on actual time, (al) establishing at least two temporalconditions, and(a2) establishing at least two spatial conditions,(b) feeding spatial conditions and temporal conditions to a data model, using artificial intelligence, and road crash data in a database, establishing an emergency, wherein artificial intelligence comprises use of a dynamic deep hybrid network,(c) based on the emergency, triaging a consequence of said road crash therewith establishing emergency of response, based on the emergency of response, dispatching an emergency vehicle to the location of the road crash or refraining therefrom.

Images