Sage Journals: Discover world-class research

Abstract

The paediatric orthopaedic expert system analyses and predicts the healing time of limb fractures in children using machine learning. As far we know, no published research on the paediatric orthopaedic expert system that predicts paediatric fracture healing time using machine learning has been published. The University Malaya Medical Centre (UMMC) offers paediatric orthopaedic data, comprises children under the age of 12 radiographs limb fractures with ages recorded from the date and time of initial trauma. SVR algorithms are used to predict and discover variables associated with fracture healing time. This study developed an expert system capable of predicting healing time, which can assist general practitioners and healthcare practitioners during treatment and follow-up. The system is available online at https://kidsfractureexpert.com/.

Keywords

expert system health Informatics machine learning paediatric orthopaedic

Introduction

A fracture is a medical condition in which the bone’s continuity is disrupted¹ affecting the bone’s cortex. When a physical force is being applied to a bone that is stronger than the bone itself, it may lead to a bone fracture incident.² Children bone’s fracture ranging from the age of 0 – 12 years old differs significantly from adults’ fracture, with 15% of skeletal trauma accounting for all injuries in children.³

A fracture that occurred in a child would take half of the time as compared to an adult for full recovery from the corresponding fracture,⁴ depending on the child’s age and fracture type. A child may take half of the time a teen usually takes to heal the same fracture type.⁵ The reason children’s fracture patterns differ from those of adults is that children’s bones are more elastic.¹ Children’s bone structure and biomechanics are different from adult bones, they will have different fracture patterns, healing mechanisms, and management as compared to adults’ fractures.³

Out of all injury cases in children, skeletal trauma accounts for 15%.³ It is vital to assess skeletal trauma since it might be able to provide insight into signalling a non-intentional injury or occurrences of unusual restoration,⁴ which will eventually lead to further discovery. For example, the child might have a medical disorder that alters the amount of time for a bone fracture to cure. Although normal adults’ bone restoration has been widely studied, the knowledge of bone fracture healing rates for paediatric cases remains sparse. It is found out that the bone fracture healing rate among paediatric is faster in comparison to adults in which study suggested that might be due to the child’s bone structure and also their youth.⁴

Conventional methods uses radiographic and x-ray fracture images combined with a statistical approach to determine the recovery time of fracture.⁵ The events leading to the injury are correlated to the fracture healing time that may be able to give some perception towards its relation to another non-accidental injury or the laceration might have different healing rates.⁴

Machine learning (ML) methods have proven that it is capable of showing a higher accuracy for diagnosis as compared to the traditional statistical methods. Various machine learning methods including Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), and Artificial Neural Network (ANN) weredat used in the adult fracture risk prediction and study of osteoporosis in postmenopausal women.^6,7 Research and studies regarding the determination of the paediatric fracture healing time are very much limited.⁸ In a previous study, ML methods such as the ANN and RF was successfully be used to predict fracture healing time in the pediatric population.^8,9

Various systems have been developed to solve specific adults orthopaedics problems, such as Osteoporosis Advisor (OPAD), the Vertebral Compression Fractures Decision Support Tool, AO PCCF Classification System and AO Surgery.^10,11,12

The Osteoporosis Advisor (OPAD) helps diagnose and treat osteoporosis. The system is designed using knowledge mapping. Expert clinicians were interviewed to determine clinically relevant criteria for osteoporosis therapy and bone mineral density (BMD) evaluation. It was intended to help clinicians identify patient risk values and provide specific diagnostic, preventative, and treatment suggestions based on global guidelines and knowledge. But the OPAD system is in Swedish and not online.¹²

Wang et al.¹¹ created an online evidence-based decision support system for differentiating benign from malignant vertebral compression fractures. The system consists of a feature checklist, a prediction model, and a reporting mechanism. Input cases are interpreted using the website's structured feature checklist. The visual gallery supports the checklist for clarity and instruction. Using the data from the checklist, a logistic regression model predicts the likelihood of malignancy. However, this work focuses on spinal compression fractures by MRI Analysis and enhancement.¹¹

Both the AO PCCF Classification System and the AO Surgery References¹⁰ have a user-friendly layout and are easily available, rich in resources, and updated with current clinical principles, procedures, and case studies. In contrast, both the system does not include a machine learning prediction element, and the users are required to determine the fracture type based on the guidelines provided.

The conventional method used in determining the type of fracture is by using the AO trauma guideline.¹⁰ The AO trauma guideline system requires the user to identify the type of fracture where orthopaedic expert knowledge is essential in identifying the right type of fracture. Therefore, an online expert system with the function that can identify fracture type and predict the healing time of bone fracture for paediatric is essential where an expert orthopaedic is not available. It is important to assist junior doctors to learn more about the treatment of paediatric fractures. It could be used by medical professionals to reassure worried parents about the length of a fracture's healing time as most cases of childhood fractures are referred to general practitioners first.

Thus, the goal of this study is to develop a web-based application system that incorporates an ML algorithm with the best performance for predicting fracture healing time and an AO trauma guideline for automatically identifying pediatric patient's fracture types and aftercare. This system can be used by medical personnel, during the course of treatment and follow-up. Doctors, orthopaedic specialists, and medical consultants, are consulted and interviewed for their suggestions, views, and requirements, which are then evaluated to arrive at a basic idea or requirement to be included in the system. In order to obtain relevant information, additional research, documentation, and a literature review are conducted. This paediatric orthopaedic expert system aims to provide medical practitioners with more information about dealing with bone fractures in children.

Materials and methods

Data collection and analysis

A collection of 242 paediatric patient data and radiographs from the paediatric orthopaedic unit department, comprising 4 years from 2009 to 2017 of infants and young children together with the time of initial injury and ages. The raw data used in this study was approved and granted permission to access the study data by paediatric orthopaedic unit. The data utilised in this study were anonymised prior to use, as our study is interested in the values and parameters without accessing patient personal information.

Type of bone fracture, segment and section of the bone involved and measurement data include the fracture angulation and fracture lengths to the physis in both anterior and lateral views are the attributes collected from the patient record examinations.¹³ The period between damage and bone union, as well as the patient's age and gender has also been determined. The time in which the bone achieved union is defined as healing time.

A total of 11 input variables are used in this study; five continuous variables and six categorical variables. The summary statistics are shown in Table 1, and for categorical variables are explained in Table 2. The healing time measured in weeks is the output variable for this study. For continuous variables, the findings are given as mean and SD, whereas for categorical variables, the results are expressed as frequencies.

Table 1.

Summary statistics of the limb data.

Variables	Attributes	Total (242)	p-Value
Age		8.7 ± 3.4	0.418
Gender	Male	196 (81.3)	0.474
	Female	45 (18.7)
Anterior diameter		39.7 ± 35.5	0.001
Anterior angulation		5.8 ± 7.9	0.581
Lateral diameter		40.3 ± 35.6	0.001
Lateral angulation		10.6 ± 13.3	0.009
Bone part	Proximal	12 (5.0)	<0.001
	Diaphyseal	88 (36.5)
	Distal	141 (58.5)
Bone segment	Epiphysis	17 (7.1)	0.003
	Metaphysis	132 (54.8)
	Diaphysis	92 (38.2)
Fracture type	Transverse	129 (53.5)	0.269
	Spiral	13 (5.4)
	Torus	69 (28.6)
	Wedge	15 (6.2)
	Greenstick	15 (6.2)
Fracture severity	Simple	235 (97.5)	0.711
	Multi-fragmentary	6 (2.5)

Table 2.

Summary of categorical data.

Variables	Division
Fracture type	1 = Transverse
	2 = Spiral
	3 = Torus
	4 = Wedge
	5 = Greenstick
Bone involved	Upper limb: 1 = Humerus; 2 = Radius/Ulna
Bone involved	Lower limb: 1 = Femur; 2 = Tibia/Fibula
Gender	1 = Male
Gender	2 = Female
Race	1 = Malay
	2 = Chinese
	3 = Indians
Bone part	1 = Proximal
	2 = Diaphyseal
	3 = Diaphysis
Bone segment	1 = Epiphysis
	2 = Metaphysis
	3 = Diaphysis
Fracture severity	1 = simple
Fracture severity	2 = Multi-fragmentary

Machine learning model development

The machine learning model development was adapted and referred from Malek’s research study.^8,14 Firstly, the raw data is normalised as some variables have a large variation or spread. The normalisation of the raw datasets is, therefore, necessary to ensure that all values of the variables are within the same range. Normalisation is essential for machine learning models^15,16.

The data was split into 70% for model training and 30% for testing. Data normalisation is performed to ensure that all values of the variables are within the same range.¹⁶ 5 – fold cross-validation is used on the training dataset as it results in a less biased or less optimistic estimate of the model performance.¹⁷ Output is then de-normalised before evaluating the model performance. The performance of the ML model is assessed using Root Mean Square Error (RMSE) to measure prediction error average level, where it calculated based on the denormalised value of the model output. The RMSE of the model quantifies the projected values and the proximity of the actual data points, demonstrating the model's ideal fit to the data.¹⁸

The ways to encounter overfitting and obtain better results are cross-validation and parameter tuning. Each model has an important parameter that cannot be detected by the data. Many parameters in the model are used to control the complexity of the model. There is also no analytical formula available to determine an appropriate value of the parameter on each model. The only method is the tuning parameter. The general approach is to define a set of values and then generate a reliable model across them.

Support Vector Regression (SVR) shows the best results in comparison to other ML algorithm such as logistic regression, random forest in Malek’s study.^8,9 The predicted variable is in numerical rather than categorical, hence SVR is applied in this study.¹⁹ Furthermore, SVR engage cost parameter to avoid over-fit of the model. The feature importances for the SVR prediction model uses Radius Basis Function (RBF) kernel and linear kernel.

Variable importance measures for the SVR model using linear kernel are obtained by building models for every predictor variable available against the response variable.²⁰ The RMSE values are then recorded and a plot of variable importance is generated. The variable which possesses the lowest RMSE is marked as the most important variable.

System development

Orthopaedic system such as OPAD and AO Trauma Surgery Reference as discussed in the introduction section was used as a guide to come up with the proposed system prototype. The paediatric orthopaedic expert system is designed to manage paediatric orthopaedic knowledge, facts and rule. The component that consists of all the experts’ knowledge is known as a knowledge base. Next, a set of actions is executed in the inference engine component if the information provided by the users fulfils the terms and conditions of the rules. The third component is the user interface, which it offers communication between non-expert users, users are required to enter data so that the logical process can start in the inference engine component.

The Expert System Architecture for the Paediatric Orthopaedic Expert System Prototype is depicted in Figure 1. The inference engine in this study consists of a trained machine learning algorithm for paediatric limb prediction algorithm adopted from the previous study by Malek’s research⁹ and expert knowledge. Expert knowledge are suggestions and knowledge gathered from the paediatric orthopaedic specialist and related documents on fracture management and healing. The inference engine executes the action whereby the user performs a query through the user interface. The knowledge base will retrieve and process data from within and pass it to the inference engine. The inference engine will start the logical process and answer the user’s query. The existing systems described in the literature are decision-support information systems that do not include a machine learning module that predicts using patient data, as in this study.

Figure 1.

System Architecture for paediatric orthopaedic expert system prototype.

Furthermore, the existing orthopaedic system is too general and does not focus on paediatric patients. This study's proposed system will include the three modules listed below: A login module that restricts system access to only registered users. This is done to ensure the security of the patient's information stored in the user account.

An update view and delete the patient’s demographic and previous data record module. A module to calculate and display the healing time of fracture of paediatric patients and to identify the fracture type with the after-guide.

To depict the flow of data in the system, a context diagram is constructed and illustrated in Figures 2 and 3 below. As the system predicts the patient fracture healing time according to the patient data provided by the approved users, which generally are general practitioners or junior doctors and trainee orthopaedics. New users are required to register and wait for approval by the admin before obtaining access to the system. Apart from identifying fracture type and predicting paediatric fracture healing time, the users can manage patient data such as adding, viewing, updating, deleting and downloading their respective patients’ data from the database in the system. The system admin can execute all the tasks accessible by the approved users and approve the newly registered users to access the system.

Figure 2.

Context diagram.

Figure 3.

SUS scores Grade Rankings from “Determining what individual SUS scores mean: Adding an adjective rating scale.”²²

System usability scale

The System Usability Scale (SUS) is used as an acceptance test for the paediatric orthopaedic expert system in this study. This method is used because it is a low-cost, quick, and reliable way to assess the usability of the system, which only consists of 10 questions.²¹ The users will rank each of the questions in the questionnaires based on the rankings shown in Table 3.

Table 3.

Scores ranking of SUS.

Strongly Disagree	Disagree	Neutral	Agree	Strongly agree
1	2	3	4	5

The scores are then converted into numbers and the usability score is calculated using SUS. The outcome of the calculated score is average at the SUS score of 70, which is considered Good according to the SUS score shown below:

SUS was chosen as a usability test for this study due to its wide adoption and quick processing time, allowing for instant feedback and comments. As a result, the collected data may be processed quickly. SUS has a wide range of applications for many programmes and systems. The SUS score is easily understood and can be improved.²²

Result and discussion

Paediatric limb fracture healing time prediction

For the paediatric lower limb model, the RMSE result falls within the range of 2.64. The feature selection process is based on the variable importance obtained by applying the Random Forest (RF) technique, which was consistent with clinical interpretation results, demonstrating the suitability of applying RF techniques in this domain. In this study, we use the mean decrease in accuracy importance measure which uses an unbiased splitting criterion and avoids both systematic bias and the increased variance. Additionally, we validate the selected features using Self-Organizing Maps (SOM) to ensure their robustness. The Sequential Backward Selection (SBS) method is utilized for the feature selection process as well.²³

From our previous study regarding the paediatric upper limb model, we used SVR and RF models to predict the upper limb fracture healing time. We applied the SBS method and variable importance to select the optimal features for the models. The selected features included gender, age, bone involved, bone part, bone segment, fracture type, fracture severity, lateral and anterior fracture distance to physis, and lateral and anterior angulation. The K-Fold cross-validation approach was employed to evaluate the models' performance. The RMSE was used to measure the models' accuracy. The final RF and SVR models with selected features outperformed the models using all features. The RMSE value of the SVR model improved from 2.62 to 2.55 after feature selection. The selected features and their corresponding RMSE values on the test set were provided for the models.⁸

Figures 4 and 5 illustrate the variable importance plot obtained from the SVR model for the variables associated with fracture healing time and plot of RMSE against reduced variables SVR respectively.^9,14 The variables' importance is ranked against fracture healing time. RMSE value recorded based on ranked variables using the SBS method is illustrated in Figure 4 of the SVR model. The Higher RMSE value recorded indicates the significance of the variable. Fracture distance and angulation, age, and the bone part where the fracture occurs are identified as important predictors to fracture healing time.

Figure 4.

A Pot of feature importance from SVR variable importance model.

Figure 5.

Sequential Backward Elimination on ranked variables based on SVR variable importance method.

Regarding the choice of machine learning algorithm used in our study, we selected the Random Forest (RF) method based on our previous research⁹ for paediatric lower limb, which reported better accuracy with RF.^24,25 While RF has been widely applied in computational biology and medicine for complex, highly correlated predictor relationships^26–28 its application in the field of paediatric orthopaedics, and specifically for the prediction of upper limb bone healing time, has not been previously reported.

In our prior research on the paediatric upper limb model,²⁹ we chose to apply the Support Vector Regression (SVR) as the machine learning algorithm for the expert system. We selected SVR because it is a non-parametric technique that does not depend on the underlying distributions of independent and dependent variables, and it is determined by the kernel function. In comparing the results of SVR with those of Random Forest (RF), we found that SVR follows the trend of healing time (measured in weeks) between the actual and predicted values. This could be due to the fact that SVR for regression is known for its ability to recognize patterns and fit functions.

Kids fracture expert system - patient management aftercare guide

This study is a developed system prototype focusing on orthopaedic paediatric, including various functions; fracture identification, bone fracture healing rate and fracture management aftercare. The system prototype uses the decision support system concept for fracture identification. The fracture healing time prediction is deployed using a machine learning algorithm. The system also integrates a patient management guide after identification of the fracture type among children’s patients, expert suggestion regarding aftercare of the fracture is also incorporated and displayed to the user.

Expert knowledge is collected and applied to the system using the decision support system concept. The illustration in Figure 6 can be found in the system for ease of user in identifying the fracture type among children.

Figure 6.

Fracture illustrations in paediatric orthopaedic system.

The graphics of paediatric trauma aftercare algorithm with expert suggestions are shown in Figure 7 where upon user inputting the types of fracture and the fracture details, the system will suggest the aftercare guide, including back slab, applying POP cast or direct reference to an orthopaedic specialist. The back slab is the simplest form of plaster splint, which is to provide support with less risk of limb constriction.

Figure 7.

Aftercare guide paediatric orthopaedic system.

Kids fracture expert system – fracture healing time and type prediction

Kids Fracture Expert is an online system that is able to predict the fracture healing time and type for paediatric patients. The machine learning model using SVR algorithm is integrated into the online expert system. It is mainly provided for medical practitioners especially orthopaedists to aid and provide insights for them regarding the expected healing time of the patient. The outcome of the expert system is shown and discussed in this section, the user interface is included and described in detail. The website URL - http://kidsfractureexpert.com/

Figure 8 depicts the homepage of the Kids Fracture Expert System, which includes a demo of the expert system that allows users to have a quick run-through by providing the required input, and the bottom of the page displays the predicted healing time and type based on the user's input information.

Figure 8.

Kids fracture expert homepage.

To access the full system, the user is required to log in to their account. For new users, they are required to register a new account using their email and set up their password for their permission to use the system. Figures 9 and 10 show the system login page and also the new user registration page.

Figure 9.

Kids fracture system user login page.

Figure 10.

Kids fracture expert new user registration page.

Once the user successfully created account and login to the system, Figure 11 below displays the expert system dashboard. The dashboard displays the patient’s name and the user can ‘Add Case’, as the patient might have more than one fracture. For the ‘View’ button, it allows the user to view the input data that has been stored in the database while the‘Edit’ button allows the user to edit that patient data and the ‘Delete’ button is used to delete the patient’s information. As for new patient, the user will need to register the patient and input the required information, thus the system can predict the healing time.

Figure 11.

Figure 4.29: Kids fracture system homepage dashboard.

Figure 12 shows the pages that allow the user to edit patient information. The ‘Case Information’ requires the user to enter information such as date of entry, date of injury, time of injury, hospital name, place of incident, the road of the incident, previous fracture, previous fracture description, medical history, the surface of impact, mechanism of injury, witness, and witness description. Then, under the ‘Bone Involved’ section, there is an interactive skeleton in which the user is able to select the bone part that is injured, followed by the segment of bone, fracture morphology, Anterior-Posterior (AP) view, lateral view, and fracture distance to physis. Once the information is done updating, the user has to click on the ‘Update’ button and the updated data will be stored in the database.

Figure 12.

Kids fracture expert patient fracture case information page.

Figure 13 shows the basic patient’s demographic information page, including name, IC number, gender, age, residence type, and parents’ information which comprises the parents’ education and occupation background Figure 14. Once the ‘Submit’ button is clicked, the information is automatically updated in the database and the updated data is displayed on the dashboard as shown in Figure 13.

Figure 13.

Patient basic information page.

Figure 14.

Result page that shows the patient information according to case and the predicted healing weeks.

The page shown in Figure 15 is the detailed view of the fracture injury along with the predicted fracture healing time and the expert suggested treatment for the injury. The images of the bone involved; the segment of bone is also displayed to the user.

Figure 15.

Kids fracture system detailed case with expected healing time and expert suggest treatment.

System usability scale test

The evaluation form is created based on the System Usability Scale (SUS) which comprises 10 questions to assess the usability and functionality of the website. The system usability evaluation form is given in the form of Google Forms to several users that are related to the research, such as;

- Paediatric orthopaedic

- General orthopaedic

- General medical practitioner

- Healthcare workers and Nurses

- Administrative staffs

- Researchers and students

The pie chart in Figure 16 illustrates the demography of the respondent designation. Healthcare workers account for 45.5% of all respondent designations; however, general orthopaedic, paediatric orthopaedic, and others account for 9.1% of the total chart.

Figure 16.

Respondent designation pie chart.

The scores are then converted into numbers and the usability score is calculated using SUS. The outcome of the calculated score is average at the SUS score of 80, which is considered “Good” and falls under the acceptability range according to the SUS. The result indicates that the system is acceptable and could be used by medical practitioners

Conclusion

This study created a useful tool for medical students and practitioners, and the research investigates the feasibility of combining machine learning algorithms such as Support Vector Regression (SVR). With a well-developed model, it can be applied to an online expert system to assist general practitioners and medical students by providing advice and suggestions for estimating the rate of fracture healing and type of fracture in situations where expert orthopaedics are unavailable.

It may be concluded that employing such a map in conjunction with fracture presentation may be a helpful screening technique in identifying children at risk of the blocked prolonged healing period, which may necessitate particular care. The fracture presentation can serve as a clinical guideline to the attending doctor as to the expected healing times of fractures. A comprehensible guideline is given as to the rate of healing and subsequent return to premorbid function of the children. plo While this research is based upon the clinical data, it can be a valuable tool for placing a patient in a clinical context, facilitating clinician consensus, assessing a patient's unique risk, and tracking their progress throughout therapy if utilised inside a validation system and periodically rebuilt as additional data is received.

However, further study is essential to improve the performance of the machine learning algorithm, and the models built through the use of SVR have not yet reached their full potential, particularly in the fracture healing rate of children. It still needs to be validated externally by using different machine learning methods such as Logistic Regression (LR), XGBoost, Random Forest (RF) and also in the unsupervised machine learning domain. Moreover, machine learning methods applications are yet to be fully developed especially in paediatric fracture healing.

Several recommendations and enhancements could be made to improve the model's performance and efficiency, including the system's usability.^30–37 As the system developed is more focused on paediatric patients, additional knowledge on predicting fracture healing time for adult patients can be added into the system in the near future, provided with the data available for adult patients. Besides, the improvement in the data imputation is essential since the data with missing values have been removed as it eventually decreases the number of training and testing datasets. And, data imputation should be applied as it will not affect any medical information obtained and it provides a large dataset for the model to train to recognise all the patterns that are able to classify groups accurately. Thus, it will improve the performance of the machine learning model.

Footnotes

Author’s note

A collection of 242 paediatric patient data and radiographs from the paediatric orthopaedic unit department,University of Malaya Medical Centre (UMMC),Wilayah Persekutuan,Malaysia,comprising 4 years from 2009 to 2017 of infants and young children together with the time of initial injury and ages. The raw data used in this study was approved and granted permission to access the study data by UMMC paediatric orthopaedic unit. The data utilised in this study were anonymised prior to usage as our study data are interested in the values and parameters without accessing patient personal information.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research,authorship,and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research,authorship,and/or publication of this article: This work was supported by the University of Malaya grant (Project No: GPF013B-2018). The funders have no role in study design,data collection and analysis,decision to publish,or preparation of the manuscript.

Ethical statement

ORCID iD

Sorayya Malek

Data availability statement

Data is available in a public repository that does not issue DOIs. The data that support the findings of this study are available from the Department of Orthopaedic Surgery,University of Malaya Medical Centre (UMMC) but restrictions apply to the availability of these data,hence,they are not publicly available. The data belongs to the individual ministry of health,universities hospitals and private hospitals that require multiple institutional agreements for data release to third parties,hence,ethical approval is needed for analysis. Data are however available from UMMC upon request using

or email to ummc@ummc.edu.my . Any findings from the data needs to be reported and permission is needed from the Department of Orthopaedic Surgery,University of Malaya Medical Centre (UMMC) committee before publication.

References

Nordqvist

Petersson

Redlund-Johnell

. Mid-clavicle fractures in adults: end result study after conservative treatment. J Orthop Trauma 1998; 12(8): 572–576.

Davis

KF.

What is orthopedics [Internet]. Medical News Today. 2020 [cited 2023 Nov 23]. Available from: https://www.medicalnewstoday.com/articles/what-is-orthopedics#definition

Staheli

. Fundamentals of pediatric orthopedics. Philadelphia: Lippincott Williams & Wilkins, 2008.

Ogden

. Skeletal injury in the child. New York: Springer Science & Business Media, 2006.

Gupta

Alderliesten

Benedictus

. A review of T-stress and its effects in fracture mechanics. Engineering fracture Mechanics. Amsterdam: Elsevier, 2015.

Burges

. A tutorial on support vector machines for pattern recognition. Data Min Knowl Discov 1998; 2(2): 121–167.

Cristianini

Shawe-Taylor

. An introduction to support vector machines and other kernel-based learning methods. Cambridge: Cambridge University Press, 2000.

Lau

Sorayya

Gunalan

, et al. Paediatric upper limb fracture healing time prediction using a machine learning approach All Life. Abingdon, UK: Taylor & Francis Group, 2022.

Malek

Gunalan

Kedija

, et al. Random forest and self organizing maps application for analysis of pediatric fracture healing time of the lower limb. Neurocomputing 2018; 272: 55–62.

10.

AO Foundation.

Orthopedic trauma: Adult trauma [Internet]. AO Surgery Reference. 2017 [cited 2023 Nov 23]. Available from: https://surgeryreference.aofoundation.org/orthopedic-trauma/adult-trauma

11.

Wang

Jeanmenne

Weber

, et al. An online evidence-based decision support system for distinguishing benign from malignant vertebral compression fractures by magnetic resonance imaging feature analysis. J Digit Imaging 2011; 24(3): 507–515.

12.

Halldorsson

Bjornsson

Gudmundsson

, et al. A clinical decision support system for the diagnosis, fracture risks and treatment of osteoporosis. Comput Math Methods Med 2015; 2015: 189769.

13.

Malone

Sauer

Fenton

. A radiographic assessment of pediatric fracture healing and time since injury. J Forensic Sci 2011; 56(5): 1123–1130.

14.

Malek

Gunalan

Kedija

, et al. Random forest and Self Organizing Maps application for analysis of pediatric fracture healing time of the lower limb. Neurocomputing. 2018 Jan 10;272:55-62

15.

Ogasawara

Martinez

De Oliveira

, et al. Adaptive normalization: a novel data normalization approach for non-stationary time series. In: The 2010 International Joint Conference on Neural Networks (IJCNN), Barcelona, 2010, 1–8. IEEE

16.

Shen

Gong

Cai

, et al. Normalization and integration of large-scale metabolomics data using support vector regression. Metabolomics 2016; 12(5): 89.

17.

Kim

Drake

Park

. Adaptive nonlinear discriminant analysis by regularized minimum squared errors. IEEE Trans Knowl Data Eng 2006; 18(5): 603–612.

18.

Armstrong . Evaluating forecasting methods. In: Principles of forecasting. Boston: Springer, 2001, pp. 443–472.

19.

Vapnik

Golowich

Smola

. Support vector method for function approximation, regression estimation and signal processing. Advances in neural information processing systems. Denver, CO, USA: MIT Press, 1997, pp. 281–287.

20.

Guyon

Elisseeff

. An introduction to variable and feature selection. Journal of machine learning research 2003; 1: 1157–1182.

21.

Brooke

. Sus: a “quick and dirty’usability Usability evaluation in industry. CRC Press: UK, 1996, p. 189.

22.

Bangor

Kortum

Miller

. Determining what individual SUS scores mean: adding an adjective rating scale. Journal of usability studies 2009; 4(3): 114–123.

23.

Malek

Gunalan

Kedija

, et al. Random forest and self organizing maps application for analysis of pediatric fracture healing time of the lower limb. Neurocomputing 2018; 272: 55–62.

24.

Breiman

. Random forests. Mach Learn 2001; 45: 5–32.

25.

Svetnik

Liaw

Tong

, et al. Random forest: a classification and regression tool for compound classification and QSAR modeling. J Chem Inf Comput Sci 2003; 43(6): 1947–1958.

26.

Cabras

Castellanos

Biino

, et al. A strategy analysis for genetic association studies with known inbreeding. BMC Genet 2011; 12: 63–72.

27.

Liu

Ackerman

Carulli

. A genome-wide screen of gene–gene interactions for rheumatoid arthritis susceptibility. Hum Genet 2011; 129: 473–485.

28.

Lunetta

Hayward

Segal

, et al. Screening large-scale association study data: exploiting interactions using random forests. BMC Genet 2004; 5: 1–13.

29.

Lau

Malek

Gunalan

, et al. Paediatric upper limb fracture healing time prediction using a machine learning approach. Life 2022; 15(1): 490–499.

30.

Budd

The basic types of pediatric fractures, differences from adults and care as a primary care physician [Internet]. 2012 [cited 2023 Nov 23]. Available from: http://learn.pediatrics.ubc.ca/body-systems/musculoskeletalsystem/

31.

Zhao

Huang

, et al. Vascularized bone grafting fixed by biodegradable magnesium screw for treating osteonecrosis of the femoral head. Biomaterials 2016; 81: 84–92.

32.

Tseng

Hung

Shieh

, et al. Hip fracture risk assessment: artificial neural network outperforms conditional logistic regression in an age-and sex-matched case control study. BMC Musculoskelet Disord 2013; 14(1): 207.

33.

Collins

Evans

. Animal: validation and applications of nonlinear registration-based segmentation. Intern J Pattern Recognit Artif Intell 1997; 11(08): 1271–1294.

34.

Kohonen

. Self-organization and associative memory. self-organization and associative memory. In Springer Series in Information Sciences. Verlag Berlin Heidelberg New York: Springer, 1988, p. 312.

35.

Hollmen

. Process modeling using the self-organizing map. Finland: Citeseerx, 1996.

36.

Geisser

Predictive inference. New York: CRC Press, 1993. vol. 55.

37.

Bhat

System Usability Scale [Internet]. 2018 [cited 2023 Nov 23 ]. Available from: https://www.questionpro.com/blog/system-usability-scale/