Proceedings of the International Conference on Augmented Intelligence and Sustainable Systems (ICAISS-2022)

IEEE Xplore Part Number: CFP22CB2-ART; ISBN: 978-1-6654-8962-1

Bone Health Prediction using Machine Learning

Approach
M anisha G. Fal Dessai Dr. Venkatanareshbabu Kuppili
Department of Computer Science and Engineering Department of Computer Science and Engineering
National Institute of Technology, Ponda, Goa National Institute of Technology, Ponda, Goa
manishafaldessai@[Link] venkatanareshbabu@[Link]

Abstract— Osteoporosis is a vertebral illness that leads to bone the surrounding soft tissues. Additionally, DXA involves
2022 International Conference on Augmented Intelligence and Sustainable Systems (ICAISS) | 978-1-6654-8962-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICAISS55157.2022.10010985

tissue loss by making bones brittle and fragile as a result of radiation exposure and more expensive equipment, which
hormonal changes, vitamin D deficiency, or calcium shortage. complicates the diagnosis. Numerous studies have been
To avoid bone fractures and their socioeconomic costs,
conducted recently to determine osteoporosis by assessing the
osteoporosis prophylaxis and early identification are crucial.
The procedure for identifying osteoporosis is time -consuming
viscosity of the collarbone cortex using thorax radiographs. A
and expensive. Due to a lack of high-tech equipment, this T-score value is used to diagnose osteoporosis, whose formula
disease is challenging to identify in rural locations. S o in this is given as,
paper, analysis of the series of classification and gradient ′
𝑃𝑎𝑡𝑖𝑒𝑛𝑡 𝑠 𝐵𝑀𝐷 𝑚𝑒𝑎𝑛 𝐵𝑀𝐷 𝑜𝑓 𝑦𝑜𝑢𝑛𝑔 𝑛𝑜𝑟𝑚𝑎𝑙 𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛
boosting algorithms like Random Forest, S VM, XGBoost, T-score=
𝑆𝐷 𝑜𝑓 𝐵𝑀𝐷 𝑜𝑓 𝑦𝑜𝑢𝑛𝑔 𝑛𝑜𝑟𝑚𝑎𝑙 𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛
AdaBoost, and CatBoost is done to predict the condition of the
bones, i.e., whether a person has osteopenia, osteoporosis, or
severe osteoporosis or has healthy bone using attributes like According to WHO criteria, the following T -score values are
age, sex, steroids taken, history of fracture, weight, height, used to diagnose osteoporosis [14]:-
smoking habits, alcohol consumption, etc. Our experiments If a person’s T-score is greater than -1 SD, then his bone
show that the gradient boosting algorithm, i.e., CatBoost, gives density is normal.
classification accuracy of 100%, precision (100%), F1-S core If a person’s T-score lies between -1 SD and -2.5 SD, then he
(100%), and recall (100%), for 75–25 splits, and surpasses the is suffering from osteopenia.
other algorithms that were considered in this paper. Table 1 If a person’s T-score is less than -2.5 SD, then he is suffering
below shows the performance comparison table of all the
from osteoporosis.
algorithms with respect to data splits.

Keywords— Prediction, classification algorith ms, The risk factors for osteoporosis are classified as majo r
osteoporosis, osteopenia, BMD (Bone Mineral Density), modifiable and major non-modifiab le. Insufficient nutrition
Random Fo rest, SVM (Support Vector Machine), XGBoost, absorption, a lack of exercise or a danger of falling, weight
AdaBoost, and CatBoost loss, cigarette smoking, and alcohol use are significant
modifiable risk factors. A history of falls, advanced age,
I. INT RODUCT ION gender, white ethnicity, a past fracture, and reproductive
variables (family history of osteoporosis) are among the main
Osteoporosis is one of the most common age-related
non-modifiable risk factors. Secondary causes of osteoporosis
problems worldwide. Th is disease affects over 200 million
people world wide, and the burden of osteoporosis is expected include chronic med ication use (prolonged corticosteroid use),
hypogonadism, hyperparathyroidis m, chronic liver dis ease,
to rise significantly in the near future, with negative
inflammatory diseases (rheumatoid arthrit is, for examp le), a
consequences in terms of bone fractures, quality of life
reductions, mortality, chronic pain, loss of autonomy, an d deficiency of v itamin D, renal disease (history of kidney
stones), cardiovascular disease, diabetes mellitus, and
social costs [2].
dementia. A lthough there is a growing awareness of the
Both men and wo men can develop osteoporosis at any age. osteoporosis illness, there are still insufficient diagnostic and
preventative measures in place [2].
Menopausal wo men have a greater risk o f developing a bone
fracture due to osteoporosis if their age is above 50 [2]. II. LITERATURE REVIEW
Because the bones affected by osteoporosis are unable to
accommodate physical pressure and pain, they crack. The hip, Several researchers have worked on the prediction of
back, and ankles are the most common sites for fractures in osteoporosis either using textual data, DXA images, or
people with osteoporosis. Different osteoporosis diagnostic radiographs. In [1], an intelligent bio-inspired method is
imaging techniques include Dual Energy X-Ray proposed that classifies normal and osteoporosis patients using
Absorptiometry (DEXA/DXA), QUS, and QCT. Osteoporosis the patient's history with 94% accuracy. Many studies have
can also be detected by magnetic resonance imaging used Gabor filters and GLCM with classificat ion algorithms to
techniques, which include MRI, diffusion-weighted imaging, extract texture features fro m X-ray and DXA images to predict
high-resolution magnetic resonance imaging, u ltra-short echo, osteoporosis [9] [10]. In some studies, BM D is predicted using
magnetic resonance spectroscopy, etc. Among these, DXA is radiographs, and then based on this, fracture risk is predicted
known as the gold standard for diagnosing osteoporosis using [3]. In [5][15] novel co mputer-based automatic detection and
bone minerals. The procedure used to confirm an osteoporosis classification methods are designed for real t ime detection of
diagnosis is to use DXA to estimate the minerals in the lu mbar disease. Furthermore, many studies have been conducted using
spine and proximal femur [6] [8] [12]. The d isadvantage of mach ine learning and deep learn ing algorithms for the early
using DXA is the presence of measurement errors caused by prediction of osteoporosis [7][11][13].

978-1-6654-8962-1/22/$31.00 ©2022 IEEE 382

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on December 09,2025 at 17:09:07 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Augmented Intelligence and Sustainable Systems (ICAISS-2022)
IEEE Xplore Part Number: CFP22CB2-ART; ISBN: 978-1-6654-8962-1

III. METHODOLOGY the percentage of results the model correctly

The proposed methodology includes 4 steps: predicted. Eq. (1) shows how to calcu late accuracy,
where P is the total number of predict ions or
A. Obtaining the dataset
guesses and Pc is the number of accurate guesses.
B. Data Preprocessing Accuracy is described in terms of positive and
C. Attributes Selection negative in the domain of binary classification, as
demonstrated in Eq. (2).
D. Implementing different ML algorithms to predict or
detect the disease. Accuracy= Eq.(1) or
The overview of the methods utilized in this paper is depicted
in the figure below.
Accuracy= Eq. (2)

Where TN True Negative,

TP True Positive,
FP False Positive and
FN False Negative
ii. Precision: A measurement called precision reveals
the proportion of patients with osteoporosis who
actually have it. Precision can be calcu lated as
shown in Eq. (3). Precision ind icates how
frequently positive outcomes were actually
achieved when something positive was predicted.

Precision = Eq. (3)

iii. Recall: Recall is a statistic that shows us what

proportion of patients with osteoporosis the
algorith m actually diagnosed as having the
condition. Recall can be calculated as Eq. (4)
illustrates. It is also referred to as sensitivity. Recall
Figure 1: Block Diagram for Bone Health Prediction indicates the number of t imes the algorithm made a
correct prediction based on actual positive data.
This paper makes use of a dataset in CSV format obtained
fro m Kaggle, namely the BM D dataset. It has nine attributes Recalll = Eq. (4)
or features, including id, age, sex, height in centimeters, iv. F1- Score: The mean of the calculated precision
weight in kilograms, fracture, medication, s moking, and
and recall values is what is referred to as the F1-
bmd. The dataset contains 169 tuples (rows), i.e., the data of
169 patients. The next step is data preprocessing, which, score value. It is more help ful when classes are
depending on the algorith ms employed, addresses missing distributed unevenly. F1- score can be computed
values, cleans up the data, and normalises it. Fo llo wing data using the formula in Eq. (5).
pre-processing, the model is then trained using different
ratios of train and test datasets, i.e., 75–25, 60– 40, and 50– F1-Score = Eq. (5)
50 splits. Machine learning algorith ms will use these
attributes or features to predict outcomes [1]. The proposed B. Experimental Results
model emp loys the following classifiers: random forest,
SVM, XGBoost, AdaBoost, and CatBoost. The developed The follo wing Machine learning algorithms were applied to
model is then put into practise, and its performance and preprocessed data. In the calculated values below, 0 is for a
accuracy are assessed using a variety of performance healthy bone condition, 1 is for osteopenia, 2 is for
metrics, including precision, recall, and F1-Score. Here in osteoporosis, and 3 is for severe osteoporosis.
this model, an effect ive bone health prediction system is
developed using different classifiers. 1. Random Forest: It is an ensemble learning technique
that builds several numbers of decision trees during
IV. RESULTS AND DISCUSSION training to do classificat ion, regression, and other tasks.
A. Evaluation Metrics For classificat ion problems, the random forest output is
i. Accuracy: One metric used to evaluate the class that the majority of the trees chose. Follo wing
classification models is accuracy. Informally, it's are the calculated values using Random Forest for the
best split (50-50).

978-1-6654-8962-1/22/$31.00 ©2022 IEEE 383

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on December 09,2025 at 17:09:07 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Augmented Intelligence and Sustainable Systems (ICAISS-2022)
IEEE Xplore Part Number: CFP22CB2-ART; ISBN: 978-1-6654-8962-1

Precision Recall F1-score Support and cases that are already well-classified less weight.
0 0.97 1.00 0.98 30 Following are the calculated values using AdaBoost for
1 1.00 0.97 0.99 36 the best split (i.e., 75–25).
2 1.00 0.83 0.91 6 Precision Recall F1-score Support
3 0.93 1.00 0.96 13 0 1.00 1.00 1.00 11
Accuracy 0.98 85 1 0.85 1.00 0.92 22
macro avg 0.97 0.95 0.96 85 2 1.00 0.00 0.00 4
weighted avg 0.98 0.98 0.98 85 3 1.00 1.00 1.00 6
Accuracy 0.91 43
Table 1: Classification Report of Random forest for 50-50
macro avg 0.96 0.75 0.73 43
split
weighted avg 0.92 0.91 0.86 43
2. SVM: It is a supervised learning technique that works
well when there are more features than data points Table 4: Classification Report of AdaBoost for 75-25 split
available. Because it uses a subset of training points
known as support vectors, its decision function is 5. CatBoost: CatBoost is a technique for decision trees
memo ry-efficient. Following are the calcu lated values that uses gradient boosting. In addition to having the
using SVM for the best split (i.e., 75-25). best AUC score for the test data (the higher the AUC
score, the higher the model's performance), it also has
Precision Recall F1-score Support the quickest prediction time without categorical
0 0.77 0.91 0.83 11 support. Following are the calcu lated values using
1 0.86 0.82 0.84 22 CatBoost for the best split (i.e., 75-25).
2 0.00 0.00 0.00 4
precision recall f1-score support
3 0.67 1.00 0.80 6
0 1.00 1.00 1.00 11
Accuracy 0.79 43
1 1.00 1.00 1.00 22
macro avg 0.57 0.68 0.62 43
2 1.00 1.00 1.00 4
weighted avg 0.73 0.79 0.75 43
3 1.00 1.00 1.00 6
Table 2: Classification Report of SVM for 75-25 split accuracy 1.00 43
macro avg 1.00 1.00 1.00 43
3. XGB oost: Ext reme gradient boosting is an ensemble weighted avg 1.00 1.00 1.00 43
mach ine learn ing approach built on decision trees that
makes use of the grad ient boosting framework. It’s a Table 5: Classification Report of XGBoost for 75-25 split
technique where additional models are used to identify
and fix mistakes caused by earlier models. The prio r
tree's prediction mistakes are used to iteratively train Sr. ML Algorithms Parameters
the model. The prev ious trees are then blended with the No
n_estimators = 100
current trees to provide the final p rojection. Follo wing
1. Random Forest i.e. The number of trees in the
are the calculated values using XGBoost for the best
forest
split (i.e., 75-25).
Precision Recall F1-score Support 2. No Parameters specified
SVM
0 1.00 1.00 1.00 11
1 1.00 1.00 1.00 22 3. No Parameters specified
XGBoost
2 1.00 1.00 1.00 4
n_estimators= 50,
3 1.00 1.00 1.00 6
learning_rate= 1,
Accuracy 1.00 43 AdaBoost
4. random_state= 0
macro avg 1.00 1.00 1.00 43
weighted avg 1.00 1.00 1.00 43 iterations= 50,
depth= 3,
Table 3: Classification Report of XGBoost for 75-25 split 5. CatBoost learning_rate= 0.1,
loss_function= 'RMSE'
4. AdaBoost: An adaptive boosting classifier is a meta -
estimator that allo ws mu ltip le "weak classifiers" to be
combined into a single "strong classifier." It works by Table 6: Parameter Specifications of all ML Algorithms.
giving cases that are difficult to categorise more weight

978-1-6654-8962-1/22/$31.00 ©2022 IEEE 384

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on December 09,2025 at 17:09:07 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Augmented Intelligence and Sustainable Systems (ICAISS-2022)
IEEE Xplore Part Number: CFP22CB2-ART; ISBN: 978-1-6654-8962-1

Metrics 1
Preci- F1 0.9
ML Accuracy Recall
Data sion Score 0.8
Algori-
Split 0.7
thms
0.6
75 - 25 0.98 0.96 0.94 0.95 0.5
Random
60 - 40 0.96 0.92 0.92 0.92 0.4
Forest
0.3
50 - 50 0.98 0.97 0.95 0.96
0.2
75 - 25 0.79 0.57 0.68 0.62 0.1
SVM 60 - 40 0.76 0.57 0.66 0.61 0
RF SVM XGBoost AdaBoost CatBoost
50 - 50 0.78 0.59 0.59 0.59
75 - 25 1.00 1.00 1.00 1.00 75 - 25 Split 60 - 40 Split 50 - 50 Split

XGBoost 60 - 40 0.97 0.93 0.93 0.93 Figure 3: Graph comparing Precision(Y) v/s Algorithms(X)
50 - 50 0.98 0.95 0.94 0.94
75 - 25 0.92 0.96 0.75 0.73
75 - 25 split 60 - 40 split 50 - 50 split
AdaBoost 60 - 40 0.90 0.93 0.72 0.70
1
50 - 50 0.92 0.91 0.74 0.70
0.8
75 - 25 1.00 1.00 1.00 1.00
CatBoost 60 - 40 0.96 0.90 0.92 0.91 0.6

50 - 50 0.98 0.95 0.95 0.95 0.4

Table 7: Performance Comparison table of all ML algorithms 0.2

with respect to data splits
0
The outcomes of all machine learning algorith ms are shown in RF SVM XGBoost AdaBoost CatBoost
Table 1 above, wh ich includes four parameters: accuracy,
precision, f1-score, and recall. These values are also obtained Figure 4: Graph comparing Recall(Y) v/s Algorithms(X)
for all the data splits. We can see that CatBoost gives the best
accuracy (100%) for a 75/25 data split. Furthermore, we can
see that accuracy, precision, f1-score, and recall values are 75 - 25 split 60 - 40 split 50 - 50 split
high for a 75–25 split for all algorith ms except for rando m 1
forest.
0.8

75 - 25 Split 60 - 40 Split 50 - 50 Split 0.6

1
0.9 0.4
0.8
0.7 0.2
0.6
0.5 0
0.4 RF SVM XGBoost AdaBoost CatBoost
0.3
0.2
0.1 Figure 5: Graph comparing F1-Score(Y) v/s Algorithms(X)
0
V. CONCLUSION
RF SVM XGBoost AdaBoost CatBoost
(Random In this paper, we attempt to determine if a person has
osteopenia, osteoporosis, severe osteoporosis, or a healthy
forest) bone state by analysing BMD datasets gathered from Kaggle.
In our study, multiple mach ine learning models were trained
Figure 2: Graph comparing Accuracy(Y) v/s Algorithms(X) and tested for the different data splits as shown in Table 1 for
various parameters like accuracy, precision, recall, and F1
score. It is observed that CatBoost gives the best accuracy of

978-1-6654-8962-1/22/$31.00 ©2022 IEEE 385

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on December 09,2025 at 17:09:07 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Augmented Intelligence and Sustainable Systems (ICAISS-2022)
IEEE Xplore Part Number: CFP22CB2-ART; ISBN: 978-1-6654-8962-1

100% for a 75-25 split. Also, it is observed that if 75% of the [7] Yi-T ing Lin, Chao-Yu Chu, Kuo-Sheng Hung, Chi-Hua Lu, Edward
data is used for training and 25% for testing, then all M. Bednarczyk, Hsiang-Yin Chen,Can machine learning predict
pharmacotherapy outcomes? An application study in osteoporosis,
machine learning algorith ms like XGBoost, AdaBoost, Computer Methods and Programs in Biomedicine,Volume 225, 2022,
CatBoost, etc. give the best classification accuracy. In the 107028, ISSN 0169-2607,
future, CatBoost can be used to predict osteoarthritis and [8] Engels A, Reber KC, Lindlbauer I, Rapp K, Büchele G, Klenk J,
osteoporosis using bone images. Also, a web-based Meid A, Becker C, König HH. Osteoporotic hip fracture prediction
application can be designed for doctors to predict from risk factors available in administrative claims data - A machine
osteoporosis in real time. learning approach. PLoS One. 2020 May 19; 15(5):e0232969. PMID:
32428007; PMCID: PMC7237034.
REFERENCES [9] F. Riaz, R. Nemati, H. Ajmal, A. Hassan, E. Edifor and R. Nawaz,
"Osteoporosis Classification Using Texture Features," 2019 IEEE
32nd International Symposium on Computer-Based Medical Systems
[1] P. KEERTHIKA , P SURESH , R. MANJULA DEVI, T .SENAP ATHIA, (CBMS), 2019, pp. 575-579, doi: 10.1109/CBMS.2019.00119.
R.P RAVEEN KUMARA , [Link] , AN INTELLIGENT BIO - INSP IRED [10] Kanthavel, R., and R. Dhaya. "Quantitative Analysis of Knee
SYSTEM FOR DETECTION AND P REDICTION OF OSTEOPOROSIS, Radiography." Journal of Electronics and Informatics 3, no. 3 (2021):
MATERIALS TODAY : P ROCEEDINGS 45(2021) 2010-2016, ELSEVIER, 167-177.
ICAMR 2019.
[11] Smets J, Shevroja E, Hügle T, Leslie WD, Hans D. Machine Learning
[2] RINALDI C, BORTOLUZZI S, AIROLDI C, LEIGHEB F, NICOLINI D, Solutions for Osteoporosis-A Review. J Bone Miner Res. 2021
RUSSOTTO S, VANHAECHT K, PANELLA M. THE EARLY DETECTION May;36(5):833-851. doi: 10.1002/jbmr.4292. Epub 2021 Apr 4.
OF OSTEOPOROSIS IN A COHORT OF HEALTHCARE WORKERS: IS PMID: 33751686.
THERE ROOM FOR A SCREENING PROGRAM? INTERNATIONAL
[12] Zellagui, S., Hivet, A., Mouss, M.E., & Hambli, R. (2021). Prediction
JOURNAL OF ENVIRONMENTAL RESEARCH AND PUBLIC HEALTH.
of proximal femur fracture risk from DXA images based on novel
2021; 18(3):1368.
fracture indexes. Computer Methods in Biomechanics and Biomedical
[3] HSIEH CI, ZHENG K, LIN C, MEI L, LU L, LI W, CHEN FP, WANG Y, Engineering: Imaging & Visualization, 9, 205 - 216.
ZHOU X, W ANG F, XIE G, XIAO J, MIAO S, KUO CF. AUTOMATED
[13] Alatas, B., Akin, E. (2005). Mining Fuzzy Classification Rules Using
BONE MINERAL DENSITY PREDICTION AND FRACTURE RISK
an Artificial Immune System with Boosting. In: Eder, J., Haav, HM.,
ASSESSMENT USING PLAIN RADIOGRAP HS VIA DEEP LEARNING . NAT
Kalja, A., Penjam, J. (eds) Advances in Databases and Information
COMMUN . 2021 SEP 16;12(1):5472. DOI: 10.1038/S41467-021-25779-
Systems. ADBIS 2005. Lecture Notes in Computer Science, vol 3631.
X. PMID: 34531406; PMCID: PMC8446034.
Springer, Berlin, Heidelberg.
[4] Periasamy K., Periasamy S., Velayutham, S. Zhang, Z. Ahmed, S.T.,
[14] Faulkner KG. The tale of the T-score: review and perspective.
Jayapalan, A. (2021) A proactive model to predict osteoporosis: An
Osteoporos Int. 2005 Apr;16(4):347-52. doi: 10.1007/s00198-004-
artificial immune system approach. Expert Systems,
1779-y. Epub 2004 Nov 23. PMID: 15565352.
10.1111/exsy.12708, ISSN 02664720, E-ISSN 14680394.
[15] Mishra, Akhilesh, Anuj Jaiswal, Lokesh Chaudhari, and Varsha
[5] Sayantan bhattacharya, Devesh Nair, Ayush Goyal, Computer Based
Bodade. "Health Record Management System–A Web-based
Automatic Detection and classification of osteoporosis in bone
Application." Journal: Journal of ISMAC December 2022, no. 4
Radiographs, 6th International Conference on Signal Processing and
(2021): 301- 313.
Integrated Networks, IEEE, 2019.
[6] Jang, R., Choi, J.H., Kim, N. et al. Prediction of osteoporosis from
simple hip radiography using deep learning algorithm. Sci
Rep 11,19997 (2021).

978-1-6654-8962-1/22/$31.00 ©2022 IEEE 386

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on December 09,2025 at 17:09:07 UTC from IEEE Xplore. Restrictions apply.