Techniques to propose explainability can also aid to increase trustworthiness, transfer-ability, confidence, fairness, or accessibility. Even still, muchAIresearch demonstrates poor usability, interpretability, and efficacy when it comes into contact with humans [1].
Different human-centered methodologies [75] can be employed to uncover users’ mental models and hence address these shortcomings (e.g., task reflection, interviews, co-design or card sorting and diagramming). Prior research with doctors using some of ML-based methods has discovered impediments to system use that are related to explainability [160,52] but go beyond it, such as "alert or click fatigue,"a lack of domain appropriate representations, or a lack of tools to counteract human biases. Conversely, dementia in older age is a major health concern. Preventive measures to prevent or delay dementia symptoms are of utmost importance. Yang et.al [175] created predictive models applying seven different machine learning (ML) techniques, then combined the findings from indi-vidual base models using a model ensemble methodology. A psychosocial wide variety of factors from multiple domains were investigated using a large nationally representative sample of older people. Findings provide new evidence on factors that are associated with dementia in later life.
(a)
(b)
(g) (f)
(c) (d)
(e)
Figure 4.2: Outline of the main idea of this research to use the mined association rules in the diagnosis of musculoskeletal injuries.(a)injured worker,(b)the occupational physi-cians examine him and, after(c)diagnosing the type of injury, (d)in workers OHHPs, enter information about the worker’s medical status as well as medical restrictions relat-ing to his workplace,(e)the information in the texts is extracted by NLP techniques,(f) stored in structured databases,(g)by association rules mining and regression methods, they find diagnosis and prognosis body part(s) injuries, which are used in the subsequent examinations by occupational physicians for medical recommendations.
This knowledge helps to Prognosis and diagnose occupational profiles more accu-rately and better determine workers’ workplaces to prevent further injury.
4.3.1 Data Collection
This research was conducted on medical data obtained from workers in the portuguese automotive industry. In this industry, musculoskeletal injuries and traumas are common because workers’ job positions are usually accompanied by frequent physical movements.
These injuries are different depending on the type of work of each workers. The most crucial body parts involved in workplaces in this industry are neck, trunk, shoulder (L1&R2), elbow (L&R), wrist (L&R), hand fingers (L&R), knee (L&R), foot (L&R). Each worker has anOHPPsin the medical department system in which the work status of the worker health is recorded over time. When each worker goes to the medical department, first the basic information, including the personnel number, the worker’s name, the area of employment, the name of the occupational technician, and the date, are registered in the system. After examining the workers and diagnosing the injury, the occupational physicians records theFWAin the worker’sOHPPs. In this way, it records one or more comments about the worker’s work style based on her/his health status in the system
1Left 2Right
so that the internal managers can do the necessary planning for the continuation of the worker’s task process. Each comments is about one of the injured parts of the body, and the occupational technicians´records one of the two labels, "MN-Must Not Use"or "SN-Should Not Use", based on the intensity of risk and the restrictions of the damage to that part. The "MN-Must Not Use"label indicates more severe conditions than the "SN-Should Not Use". We consideredOHPPs associated withFWAregistered in the health system from January 2019 to October 2021 for this study. The automotive industry has different areas of work, each of which has specific tasks. For this reason, workers’ job demands are different based on their workload in theOHPPsin each area. There are between 2025 and 7857 records in the mentioned period, after cleaning data, which is the range of [391, 1514] records suitable for our research. The distribution of these records among the areas of this automotive company is as follows: Assembly, Body Construction, Special Projects, Quality Assurance, Paint, and Metal Stamping. Other areas such as Special Cars, Product Management & Planning, Logistic, Plant Manager, and Finance contain less than 2% of the records from which reliable knowledge can not be extracted. This range and distribution are based on the two scenarios that will explain in section 4.3.2.
The output of the mentioned system for each record is in apdf file. a separate ap-plication designed by using the PyPDF23package provided in the Python language to extract these files’ contents automatically. The content fetched from the application is not clean data. This can be due to possible errors when entering information in the system or when converting pdf to text. These errors included non-word symbols, extra spaces, misspellings, and incorrect use of Capital letters (introducedNLPtools in section 3.4, table 3.1). That is why we used the Natural Language Toolkit (NLTK)4to clean up data, especially the data on occupational physicians’FWAstatus of workers.NLPalso was used as techniques to process each cleaned comment and extract the keywords for the "body part"and "risk label"assigned to it by the occupational physicians. Finally, the obtained information is stored in the form of tabular structured data. Each row of the final table is a record of worker visits, and each of the 14 variables in this table represents a body part.
Each item in the table indicates the type of risk for a worker’s body part (MN-Must Not use or SN-Should Not use). Table4.1shows two rows of final tabular structured data, as an example.
4.3.2 Case Studies
In Section4.3.1, it is explained that in the portuguese automotive industry, the medical team, after examining the worker’s health, diagnoses the damage and categorizes its re-strictions into two categories: MN-Must Not Use and SN-Should Not Use. Occupational technicians then record these cases in theOHPPs, along with recommendations for con-tinuing the worker’s working conditions. It is clear that asymptomatic or future injuries
3https://pypi.org/project/PyPDF2/
4https://pypi.org/project/nltk/
Table 4.1: Two rows of final tabular structured data.
medical restrictions Neck Trunk ShoulderL ShoulderR ElbowL ElbowR WristL WristR FingersL FingersR KneeL KneeR FootL FootR
1
Should not per-form tasks that involve move-ments of the right elbow.
MN-MustNotUse MN-MustNotUse SN-ShouldNotUse Must not perform
tasks that involve movements above both shoulders.
2
Must not perform tasks that involve movements of the left wrist and left fingers.
MN-MustNotUse MN-MustNotUse MN-MustNotUse Must not perform
tasks that require force with fingers of both hands.
are considered in this process as well. Below, the two contexts describe the ways in which different systems are represented according to the collected data. The first is the diag-nosis procedure, which includes a knowledge extraction system for detecting patterns between injured body part(s). The other context is the prognosis process, which predicts the severity of the next injured body part(s) in parallel to the next medical appointments.
4.3.2.1 Diagnosis procedure
The diagnosis process refers to discovering latent musculoskeletal injuries in workers using theARsmining approach.AImethods used to extract patterns in workers’ health data inOHPPsand proposed a new approach. Our proposed approach is based onARs mining. The idea is to explore the damage to the body part(s) in each specific work area by extracting ARs. For example, if in area X one of the extracted rules shows
"W ristL, ShoulderL → ElbowL", it means in that area with injury to the left wrist and left shoulder, most likely the left elbow limitation is also involved. We know that oc-cupational technicians in this industry, in addition to diagnosing the injury, also pay attention to the restrictions of the injury MN-Must Not Use or SN-Should Not Use. We also modified the final tabular dataset we generated in (4.3.1) to use it in theARsmining process. In the table, we divided each column of body parts into two columns; one for the MN-Must Not Use label and one for the SN-Should Not Use label. For example, the ShoulderLcolumn has been converted to two columns namedShoulderL−MN−MustN otU se
andShoulderL−SN−ShouldN otU se. In this way, we had a table with 28 columns, the cells inside of which filled with 0 and 1. There are many algorithms for miningARs, such as the Apriori [7], Eclat [117] and FP-Growth tree [68] algorithms. In this study we use the Apriori algorithm introduced by Agrawal et al. in 1994. This algorithm is the most classi-cal algorithm for mining frequent itemsets andARs[148]. This algorithm used to mine the ARsamong the injuries inflicted on workers in every work area of the automotive industry. The resulting rules can be provided to the medical department as additional knowledge to help diagnose injury and determineMRs. Details of this algorithm was shown in section 2.4. For example, a worker’s left hand may be injured but its symp-toms are hidden at the time of the examination, or the worker’s task at the workplace is such that the worker will soon have a back injury but have no injury at the time of ex-amination. Workers working in a particular area of the automotive industry often suffer similar injuries because their work is usually similar and in line with a specific goal. By extracting and analyzing the patterns in the results of previous examinations of workers, it can help occupational technicians to detect hidden injuries during the examination more accurately (detailed information next chapter5).
There are 2025 records betweeen 2019 to 2020, after cleaning data, which is 391 records suitable for our research. The distribution of these 391 records among the areas of this automotive company is as follows: 51% in Assembly, 20% in Body Construction, 13% in Special Projects, 6% in Quality Assurance, 4% in Paint, and 4% in Metal Stamping.
As mentioned before other areas such as Special Cars, Product Management & Planning, Logistic, Plant Manager, and Finance contain less than 2% of the records from which reliable knowledge can not be extracted. Therefore, we have not considered these areas for this research. Figure4.3shows this distribution.
Assembly 51%
Body Construction 20%
Special Projects 13%
Quality Assurance 6%
Metal Stamping 4%
Paint
4% Other
2%
Figure 4.3: Distribution of data records among the areas.
4.3.2.2 Prognosis Procedure
In this section, 7857OHPPswere based on theFWAstatus of workers. The preprocessing procedure was based on the description in sections 4.3.2 and4.3.1 was targeted. So, 1514 records were achieved after cleaning the data. These records were associated with 638 workers, which means 23.3% and 77.7% ones are female and male, respectively.
Therefore, males have a higher rate of medical appointments. Table 4.2 describes the seniority portion of workers according to Male and female. And also mean andStandard Deviation (SD)are presented based on Male and female.
Table 4.2: Seniority based on M:Male and F:Female and their min, max, mean and SD.
Gender Max Min Mean SD Length of Employment
F= 154 M=484
27 29
4 4
13.95 18.62
7.78 8.33
Among theseOHPPs, the histogram is illustrated below as a statistical summary of 638 workers based on medical appointments. Figure 4.4 illustrates the categories of workers based on medical appointment times. Hence, 220, 172, 131, 57, 32, 15, 10, 1, workers had visits to the medical department once, twice, three, four, five, six, seven, nine, and eight times. There were no workers in eight medical appointments.
0 50 100 150 200
250 220
172
131
57 32
15 10
0
1
Number of workers
Number of Appointments
1 2 3 4 5 6 7 8 9
Figure 4.4: Phases of the methodical process used in the investigation.
There is an unambiguous relationship between the number of medical appointments and the number of workers working in different areas. Figure4.5demonstrates the por-tions of records based on Assembly, Body Construction, Special Projects, Quality Assur-ance, Metal Stamping and Paint. The distribution of these 1514 records among the areas of this automotive company is as follows: 55% in Assembly, 17% in Body Construction, 5% in Special Projects, 6% in Quality Assurance, 7% in Paint, and 8% in Metal Stamping.
As mentioned before other areas such as Special Cars, Product Management & Planning, Logistic, Plant Manager, and Finance contain less than 2% of the records from which reliable knowledge can not be extracted. Therefore, we have not considered these areas for this research.
Figure 4.5: >7k OHPP between 2019 to 2021 in Assembly, Body construction, Special projects, Quality assurance, Metal stamping and paint.
A suitable, well-structured method is essential to performing sound empirical re-search. Empirical research methods are a class of research methods in which empirical observations or data are collected in order to test a theory [42]. In the present study, a quantitative method serves to test this theory of whether utilizing worker´sFWA param-eters can improve the performance ofMLapproaches inOHPPs. This section describes the four phase methodology applied in this investigation (Figure 4.6). The first prog-nosis phase centers on a literature review (section3.3) and study of preceding research work onAI. This phase has three stages: studying the academic occupational profiles such as workers or patients, studying theAIalgorithms, and studying theAIevaluation methods. With reference to the principal research objective that deals with improvingAI algorithms in academic medical investigation, the most popular academicML-NLPwas first analyzed comprehensively. Ten articles in this domain were briefly examined from varying points of view in machine learning-based natural language processing regarding
medical database (3.4). This component of the literature review has led to generating a list of general and specific features of suchML-NLP-based research, as mentioned in Chapter 3. The next stage of the first phase contains a detailed description of the most important physical key boards in FWAstatues (described in section4.3.1). The second section is devoted to providing a point out of theOHPPs text. About the database de-scribed in section 1.2, workers’ FWA status is used to create OHPPs. In other words, worker’s health protection departed from three factors that are criteria of analyzingAI algorithms. These metrics fall into three main groups based on what exactly they belong to: (1) work related parameters; (2) individual variables, and (3) organizational criteria.
HCXAIsupporting decision-making, aims to go from worker’sFWAto explanations by integrating explainability into MRs AIand supporting occupational physicians in two decision contexts: prognosis of individual, work related and organizational parameters.
The selection of relevant predictor variables was done using expert-driven methodology in the automotive industry. And, regarding this kind of selection of relevant predictor variables was crucial during the construction of regression models by pattern recognition because not all the variables contained information directly lakedFWAassociated with musculoskeletal disorders and some acted as distracting variables that only impaired the quality of the results. The first phase is related toMRsbased on eight body regions in por-tuguese texts. Physical loading (external) from outside the body produces physiological effects within the body. Individual capacity refers to the traits that define how well body tissues can withstand external stimulus and how well they respond physiologically to it.
If the biomechanical forces are too great, direct tissue damage will occur. The capacity is determined by;
• Body type and size: Strong and large people can withstand heavier loads than weak and little people.
• Gender:Women’s maximal muscle strength is around two-thirds that of men, re-gardless of body size [37].
• Age: Muscle strength increases during adolescence but begins to decline before the age of 30. This loss begins slowly, but it accelerates with age, reaching an average of 8–16 percent every decade after 50 years of age (in this study age is excluded).
• General health: Several medical disorders can weaken tissues and cause injuries to take longer to heal.
• Skills: Skilled persons can manipulate their bodies and external forces to keep biomechanical forces within the body at a minimum. Unskilled people are more likely to be involved in situations that result in accidental injuries (e.g. when losing their balance). Seniority and function are two factors in this category.
The forth section introduces the criteria of analyzing regression algorithms. The prediction of next medical appointment and next injured body part(s) was treated as a re-gression problem, where the expected outcome was continuous value. Several rere-gression methods were considered. According to the type of data collected and the final objective of the study, including Grid search and Bayesian optimizations,RFregression,ANNs,GB andGBDT. The regression methods were compared to identify the one with the highest predictive level. These metrics fall into two main groups based on what exactly they measure: measuring the rating predictionR Square (R2), usage prediction errorRMSLE (RMSLE), Mean Absolute Error (MAE), andSHAP evaluation inXAI. all of which are discussed in the last section of section 4.4. Figure4.6 shows an overview of prognosis part.
Body parts with less functional work ability Neck, Shoulder, Elbow, Knee, Trunk, Foot, Fingers, Wrist
Occupational Health Protection Profiles (text) Phase 1: Portuguese text database
Identification of variables Individual
Organizational Work-related
Measurement or evaluation of identified variable
Direct Observation Individual
Organizational
Work-related Characteristics data
FWA status
FWA status Preparation of the data
Selection of Variables Categorization
Pattern recognition techniques
Models’ construction Evaluation of models Gradient Boosting Decision tree
Decision tree Gradient Boosting
Random Forest
R2 RMSLE
MAE Phase 2: Clustering text
Phase 3: Parameter’s relationship
Phase 4: Prediction
Figure 4.6: Phases of the methodical process used in the investigation.
Regarding prediction next medical appointment, It is worthwhile to know this study devised to four main phases:
• phase 1: portugease text database the database updated in automotive industry weekly.
• phase 2: clustring textusingNLPto extract indivitual, organizational and work-relate factors by identification of appropriate varibles.
• phase 3:parameter´s relationshipusing pattern recognition for finding relation-ship with each other byARs.
• phase 4:predictionmodel´s construction with regressor models such asGBDT,DT, GBandRFby error evaluation such asR2,RMSLEandMAE
Therefore, in this scenario, for prediction process, knowing about the interval between one medical appointment to next one is significant based on how many workers have medical appointment per weeks. As figure4.7presents the number of the records based on weekly appointments. Interestingly, the mode of this distortion is 47 workers in the second week on a maximum period of 126 weeks. Over time, the rate of appointment regularly decreases.
Figure 4.7: Distribution of records based on week and medical appointment times.
4.3.2.3 Scoring methodology for work Ability Index
Table4.3provides score forWork Ability Index (WAI)based on Neupane’s (2011) multi-site pain regarding work ability among an industrial population, whether the number of musculoskeletal pain sites predicts future poor work ability [112].WAIwas measured as a subjective comparison of current work ability to a worker’s self-identified life time. The WAIwas created at the Finnish Institute of Occupational Health and tested using clinical data. TheWAIis a tool that is used in clinical occupational health care and researched in a number of countries (it has been translated into 26 languages). The index is calculated using responses to a series of questions about work demands, worker health, and resource availability. Scores range from 0 (unable to work) to 10 (maximum work ability), with excellent (scoring 10) being the highest, good (score 9), moderate (score 8) being the next highest, and poor (scores 0–7) being the lowest. Work capability is divided into two
categories: good (8–10) and poor (0–7). In this thesis, due to injuries of workers assigned to medical appointments, the range of (0-7) was took into account. So, according to table 4.4in built based on4.1, each injured body parts weighted based on severity of MN_used and SN_used. For example, one of the workers had the medical restrictions like below:
• must not perform tasks that involve performing movements above the shoulder line both
• should not perform tasks that involve performing flexion/rotation movements of the trunk
• must not perform tasks that imply performing tasks using tools with associated vibration
• should not perform tasks that involve performing left and right wrist rotation move-ments
• must not perform tasks that involve performing tasks that require force with appli-cation point on the fingers of both hands
Categorization for each medical restriction starts with SN_used =1 , MN-Must Not Use= 0.5. Scoring MN-Must Not Use: We had 2 MN-Must Not Use. So, the scoring point for 2 MN-Must Not Use is 2. We have to divided this 2 between 2 sentences and their body injured parts. In this case, 0.5 goes to shoulder_R, shoulder_L and also finger_L and finger_R. We have two SN-Should Not Use which are 5 score to injured body parts.
So, 2.5 should divided to first sentences and second one. We have to score to trunk 1.66 and wrist_R and wrist_L, 0.8. So the scoring is presented in table 4.3:
Table 4.3: Scoring to each body region(s) injuries.
trunk Shoulder_L Shoulder_R Finger_L Finger_R Wrist_R Wrist_L
worker 1.66 0.5 0.5 0.5 0.5 0.8 0.8