استخراج سناریوهای مکرر از متون روایی با استفاده از شبکه های بیزی:کاربرد برای حوادث جدی ناشی از کار با اختلال حرکت

A probabilistic approach has been developed to extract recurrent serious Occupational Accident with Movement Disturbance (OAMD) scenarios from narrative texts within a prevention framework. Relevant data extracted from 143 accounts was initially coded as logical combinations of generic accident factors. A Bayesian Network (BN)-based model was then built for OAMDs using these data and expert knowledge. A data clustering process was subsequently performed to group the OAMDs into similar classes from generic factor occurrence and pattern standpoints. Finally, the Most Probable Explanation (MPE) was evaluated and identified as the associated recurrent scenario for each class. Using this approach, 8 scenarios were extracted to describe 143 OAMDs in the construction and metallurgy sectors. Their recurrent nature is discussed. Probable generic factor combinations provide a fair representation of particularly serious OAMDs, as described in narrative texts. This work represents a real contribution to raising company awareness of the variety of circumstances, in which these accidents occur, to progressing in the prevention of such accidents and to developing an analysis framework dedicated to this kind of accident.

Prevention of trips, collisions, slips and other movement disturbances in the workplace represents an undeniable human and financial challenge. Bureau of Labor Statistics (BLS, 2012) data show that, in the USA, these accidents represented about 30% of the 1,181,290 non-fatal occupational accidents (OA) with days lost in 2011. In its 2008 document on the “causes and circumstances of accidents at work in the European Union”, the EC states that, among the 3,983,881 non-fatal accidents causing more than 3 lost work days in 2005, 19% were slips, trips, missteps, stumbles without a fall or with a fall on the level (CE, 2008). At companies operating under the French general social security system, slips, trips and other movement disturbances in work situations (excluding working at height) represented 32% of accidents with days lost (213,940 accidents); 34% of accidents with permanent partial disability (13,759 accidents); 35% of lost work days due to temporary disability (13,591,652 days) and 5% of fatal accidents (25 accidents) in 2011 (CNAMTS, 2012). Analysis of this kind of accident is often limited to factors close to the injury in the accident genesis. However, accident analysis has also revealed explanatory factors distant from the injury, such as equipment usage (Kines, 2003), access system configuration (Leclercq et al., 2007), work system design (Derosier et al., 2008), work organization or safety management (Bentley and Haslam, 2001). Each factor revealed by analyzing an accident is required for its occurrence, irrespective of its position in the accident genesis. Investigating the accident genesis as far upstream of the injury as possible therefore assists prevention in terms of highlighting a maximum number and a variety of levers for action. A diversity of OAMD occurrence circumstances for different activity sectors (Leclercq and Tissot, 2004) has also been observed within a single company (Leclercq and Thouy, 2004). Combinations of factors common to several slips, collisions and other movement disturbances have been empirically identified in all the accidents subject to in-depth analysis at a regional power distribution facility (Leclercq and Thouy, 2004) and at a railroad company (Leclercq et al., 2007). The authors termed each of these combinations a “recurrent scenario”. Haslam and Bentley (1999) had already observed that a combination of slippery conditions, use of footwear with worn treads and time-saving behavior was encountered in 50% of slip, trip and fall accidents among postal delivery workers. Representing accidents by combinations of factors, rather than by isolated factors, allows us to characterize more closely accident-causing situations since an isolated accident factor (congested floor, person running, etc.) is more representative of a usual occupational situation than of an accident. Prevention is thus more a question of controlling factors, whose combination can be harmful, rather than trying to eliminate every risk factor (Monteau, 1997), which would indeed appear illusory in the case of OAMDs. Furthermore, a fact that has contributed to accident occurrence can sometimes only be considered an accident factor within the context of its occurrence. It may, in fact, be a safety-related factor in another context. For example, knowledge of a location is a safety-related factor, when a person anticipates a step (abrupt change in level) at a location where it is unusual (e.g. midway along a corridor). This same knowledge can be an unsafeness-related factor, when there is an unfamiliar obstruction and a person, trusting his/her knowledge of the location, does not notice it. Characterizing an accident by a combination of factors, such as an accident scenario, rather than by an isolated factor therefore allows us to consider contextual information reflecting the accident-causing nature of certain identified factors. The purpose of this research is to develop serious recurrent OAMD scenarios, which go beyond the exclusively empirical stage of this development process adopted by Leclercq et al. (2007). Our work falls within the framework of a systemic accident model (Hollnagel, 2004), which has proved beyond any doubt its value to OAMD prevention (Bentley, 2009). Bayesian Network (BN)-based approaches appear better suited to answering this kind of issue. They provide an adequate representation of our pre-processed data, being a set of accident factors combinations built by experts. Each combination is composed of qualitative knowledge (accident factors) and logical links (links between factors in each logical combination). BNs are well adapted to model such data, bridging the gap between different types of knowledge and unifying all available knowledge into a single type of representation. They are capable of apprehending qualitative knowledge, in terms of accident factors, and links through BN structure. They can also apprehend quantitative knowledge, in terms of frequency of accident factor occurrence among data set, through BN parameters, allowing recurrent scenarios extraction. Unlike other methods such as neural network models, regression methods etc., all the parameters in Bayesian networks have an understandable semantic interpretation. This method can therefore combine expert knowledge with data, to build the model. This is particularly useful when the amount of data is small. Moreover, if machine learning techniques are used (with or without expert knowledge) to build the model from a data set, it can be explained in terms that are understandable by domain experts. BN-based occupational safety studies have been conducted by several authors in recent years. Using coded data, they have analyzed the effect of task performance-related factors in situations involving risks of falling from ladders or equipment such as scaffolding (Martín et al., 2009), the effect of safety climate- and individual experience-related factors on human behavior (Zhou et al., 2008) or the effect of accident factors (Zhao et al., 2012) or working conditions (García-Herrero et al., 2012) on accident occurrence. In the field of road accidents, BNs are increasingly used, e.g. to model and classify accidents according to their injury severity (Simoncic, 2004 and Oña et al., 2011) or to predict the number of accidents of different severity (Deublein et al., 2013) or crash in real time (Hossain and Muromachi, 2012). To our knowledge, no BN-based research has investigated a methodology for determining recurrent scenarios as a diagnostic step toward improving occupational safety. This aim requires in-depth analysis of a set of accidents, which can be found in a database whose richest information is contained in narrative texts. Indeed, Lincoln et al. (2004) have shown that narrative text analysis is a useful supplement to traditional epidemiological analyses because it provides qualitative data, usually based on the accident/injury process, which offers a deeper understanding of the underlying accident process. Fatality investigation reports, in particular, contain data elements not routinely analyzed with coded occupational injury surveillance data (Bunn et al., 2008). The issue now is, “Is it possible to extract recurrent scenarios from a set of serious OAMD narrative texts?” Further studies aimed at understanding accidents based on narrative text have been conducted in recent years. McKenzie et al. (2010) describe recent advances in using this kind of text in injury surveillance research. Narrative texts need to be pre-processed, unlike coded data which can be directly applied within the scope of BN-based approaches. Automatic methods, such as text mining, have been developed to extract clusters of words with a high probability of target category association (Brooks, 2008). However, these methods do not allow accurate identification of accident factors from a narrative text, i.e. facts that make sense in terms of the accident progression. Similarities or identities can effectively be expressed in words with different meanings or, conversely, a similar meaning can be expressed in differently spelt words (McKenzie et al., 2010). These facts can only be extracted, if the whole narrative is considered. To date, most OA analyses based on narrative text have implemented, for example a ‘reconstruction template’ (Lincoln et al., 2004), a priori-defined generic accident factors (Shibuya et al., 2010) or Haddon's matrix (Bunn et al., 2008) to process information manually. Analysis of these processed data is usually based on the occurrence and co-occurrence of factors or a number of related keywords. Our aim is to extract combinations of factors common to several accidents, so we need to identify, from narrative texts, both accident generic factors, which have contributed to injury occurrence and how these factors have combined to cause injury. A BN-based approach has been developed to extract such combinations or recurrent scenarios.

This paper describes a probabilistic approach to extracting recurrent scenarios from serious Occupational Accidents with Movement Disturbance (OAMDs). Analysis of this type of accident is often limited to factors close to the injury in the accident genesis. However, the causality of such accidents, indeed of any occupational accident, originates in a specific production activity context. A fundamental notion in this study is the representation of accidents by combinations of factors, rather than by isolated factors in order to characterize more closely accident-causing situations. Such combinations, common to several accidents, have been empirically identified at companies and are considered to be recurrent scenarios. In relation to preventing frequent accidents involving all workers, without exception, the mere fact that several factors contribute to the occurrence of the vast majority of such accidents is insufficient for characterizing risk situations. Their reconstitution, in the form of recurrent scenarios, is essential to more effective prevention. The proposed approach focuses on identifying factor combinations common to several accidents extracted from narrative text. It comprises four parts: a necessary initial coding step to extract relevant information from text data. Each accident narrative text is coded by logically combining generic factors inspired by the INRS model (Monteau, 1997). It should be noted that this coding step is very time consuming for experts and could not be performed on a large data set. A Bayesian Network (BN)-based model is then built for OAMDs using expert knowledge and data extracted from 143 narrative texts, which combine qualitative and quantitative aspects of the relevant knowledge. Following these initial steps, the key step in our work involves clustering OAMDs into “similarity” classes taking into account both the generic factors occurrence and pattern. Finally, the Most Probable Explanation (MPE) is derived for each cluster. Expert knowledge is therefore the essential foundation of the two initial steps in order to make up the small sample used. Consequently, validity of the identified scenarios could be confirmed by adopting BN structure learning process. This would require the use of many more cases of serious OAMDs. Moreover, these cases should be characterized not only by factors describing the injuries but also by factors explaining movement disturbances. In a context where these accidents are rarely analyzed in depth, this is a real difficulty. We successfully used a BN to represent OAMDs and extract OAMDs scenarios. Scenario richness depends on the depth of analysis, the uniformity of circumstances, in which accidents occur, and the level of generality of the generic factors. The results of our study are useful in the prevention field on the one hand for generating company awareness of the variety of circumstances, in which these accidents occur, and, on the other hand, for developing an analysis framework dedicated to this type of accident. The analysis levels defined here could be useful for developing such a framework which could also be used to get homogeneous data on OAMDs and so more precise scenarios. The present work allowed us to answer the issue stated in the introduction “Is it possible to extract recurrent scenarios from a set of serious OAMD narrative texts”? Two occupational sectors have been considered here, based on the hypothesis according to which generic factors and genesis of this kind of accidents may be different in different activity sectors. The results support this hypothesis. Nevertheless, they may be even less generalizable to all OAMDs that only particularly serious OAMDs occurred in two sectors have been analyzed. In the future, logical combinations of generic factors derived from accidents analyzed more deeply and containing less missing values, may enable the dynamic BN to be used to obtain a more accurate, comprehensive representation of OAMD genesis based on the model's dynamic aspect. This characteristic allows us to repeat the same structure in different slices, when OAMD representation may be considered in levels and slices. In other words, these slices could enable us to consider sequences of factors (or of conjunctions of factors) instead of just one factor (or conjunction of factors) at the same level of the OAMD model. Moreover, other improvement could be considered in the process such as the cross-validation for the assessment of the model accuracy. Lastly, this study takes into account the factors, considered by three experts as having had a role in accidents occurrence. What about the strength of the link between them? Causal relationships are indisputable between some of them, especially technical factors. For example, there is no doubt a slippery floor caused a slipping when accident analysis revealed a link between these two facts. However a slippery floor is not enough to lead to a slipping. Individual factors such as tiredness, experience or organizational factors leading to precipitation for example can combine with a slippery floor to cause the slipping. Even if most often technical component cannot be the only cause of movement disturbance, it is not possible to offer evidence for causal relationships among factors when individual or organizational factors are involved. The idea here is to search for recurrent combinations of factors of different nature. More often the involvement of individual and organizational factors in such combinations, more strong the “causal” link between them and other factors.