• Sonuç bulunamadı

Applying An Artificial Neural Network To Predict Osteoporosis In The Elderly

N/A
N/A
Protected

Academic year: 2021

Share "Applying An Artificial Neural Network To Predict Osteoporosis In The Elderly"

Copied!
6
0
0

Yükleniyor.... (view fulltext now)

Tam metin

(1)

Applying an Artificial Neural Network to

Predict Osteoporosis in the Elderly

Jainn-Shiun Chiua, Yu-Chuan Lib, Fu-Chiu Yuc and Yuh-Feng Wanga,1

a

Department of Nuclear Medicine, Dalin Tzu Chi General Hospital, Chiayi, Taiwan b

Graduate Institute of Medical Informatics, Taipei Medical University, Taipei, Taiwan c

Medical Intelligence Network, Taipei, Taiwan

Abstract. Osteoporosis is an essential index of health and economics in every country. Recognizing asymptomatic elderly population with high risks of osteoporosis remains a difficult challenge. For this purpose, we developed and validated an artificial neural network (ANN) to identify the osteoporotic subjects in the elderly. The study population consisted of 1403 elderly adults (mean age 63.50 ± 0.24 years ranged from 50 to 91 years old, 157 male and 1246 female) randomly selected into 3 sets, 703 participants in training set, 350 participants in selection set, and the remaining 350 participants in test set. The input variables included demographic characteristics, anthropometric measurements, and clinical data. The outcome variable was dichotomous, either non-osteoporotic (T-score of greater than -2.5) or osteoporotic (T-score of -2.5 or less) groups classified by the measurement from dual energy X-ray absorptiometry. ANN was constructed with data from training and selection sets and validated in test set whose outcome variable was unknown to the network. The performance of ANN was evaluated by discrimination and calibration simultaneously. After training processes, the final best ANN was a multilayer perceptron network which determined seven input variables (gender, age, weight, height, body mass index, postmenopausal status, and coffee consumption) as significant features. The discriminatory power of ANN for test set was excellent (area under receiver operating characteristics curve = 0.82 ± 0.03). ANN also had statistically good fit represented by statistically insignificant Hosmer-Lemeshow statistic (p = 0.24). These results suggested that our final ANN concurrently had good discriminatory power and good-fit calibration. ANN can be used as a promising tool for the elderly to stratify high risk subjects into osteoporotic group. Keywords: Neural network, Osteoporosis, Elderly.

1. Introduction

Osteoporosis is not only a healthy issue but also an economical index in every country. Recognizing asymptomatic people with high risks of osteoporosis remains a difficult challenge. The World Health Organization (WHO) has established the criteria for the diagnosis of osteoporosis based on the measurement of bone mineral density (BMD) by the dual energy X-ray absorptiometry (DEXA). Therefore, measuring BMD has been acknowledged as a useful method to predict the risk of a future osteoporotic fracture in an individual [1]. However, routine BMD measurement of all population by DEXA for screening is limited in most of the countries since the procedure is costly and the

1 Corresponding Author: Yuh-Feng Wang, MD. Department of Nuclear Medicine, Dalin Tzu Chi General Hospital. No.2, Minsheng Rd., Dalin Township, Chiayi County 622, Taiwan. E-mail: yuhfeng@gmail.com A. Hasman et al. (Eds.)

IOS Press, 2006

(2)

equipment of DEXA is not universally available. To surmount this problem, several predictive algorithms have been developed and validated for screening or monitoring.

Artificial neural network (ANN) is a computational model, being composed of parallelly nonlinear processing elements arranged in highly interconnected networks with a formation that emulates complex human thought processes such as adaptive learning, optimization, reasoning and decision making [2]. Since ANN is the most widely depicted application of artificial intelligence in clinical medicine, use of ANN to predict osteoporosis is promising. Herein, our study is the first investigation to develop an ANN model in predicting osteoporosis for the elderly and we have validated its feasibility in comparison with the measurement by using DEXA as reference method.

2. Material and Methods

2.1. Subjects

The Institutional Review Boards for Human Investigation at our hospital approved the study protocol. Subjects in this study were 2043 adults who underwent bone densitometry from 2001 to 2005. Subjects were excluded if age less than 50 years, presence of acute illness, taking medications (except steroids) known to affect bone metabolism, or any reason influencing the measurement of BMD. The final study participants consisted of 1403 elderly adults (157 male and 1246 female).

2.2. Measurements

Demographic characteristics (gender, age), anthropometric measurements (weight, height, body mass index [BMI]), and clinical data (postmenopausal status, steroid use, current smoking, alcohol drinking, coffee consumption) for all subjects before densitometry were recorded. Only one technologist was assigned to measure BMD using DEXA. All densitometrical results were interpreted by a nuclear medicine physician.

2.3. ANN construction

STATISTICA 7.0 (StatSoft, Inc., Tulsa, OK, USA) was used to generate various formulations of ANN models using a training set of subjects selected at random from the original cohort of subjects (n = 1403). We used the standard approach of cross validation to randomly separate the dataset from all participants into three parts: a training set (n = 703), a selection set (n = 350), and a test set (n = 350). Although the training set is optimized to fit ANN model, the selection set is separately used to estimate prediction error for model selection and stop training to mitigate over-learning and over-fitting. A third set known as the test set is used to validate the performance and the generation error of the final chosen model.

To choose the most excellent neural network, the automatic network designer was employed to decide an appropriate architecture. All variables were entered as continuous or nominal input variables into ANN models. The status of BMD according to the measurement of DEXA, categorized as osteoporotic (T-score of -2.5 or less) and non-osteoporotic (T-score of greater than -2.5) groups based on WHO criteria, was entered as a dichotomous output variable. During the processing procedure, the

(3)

intelligent problem solver guided a large number of experiments, which were used to determine the best architecture. It could allow simultaneous comparison of different types of networks using a combination of heuristic and optimal algorithms and automatically selected the smoothing factor and the number of units for these networks. To compare the performance of networks with different input variables, the intelligent problem solver balanced error against type and diversity as criteria to select retained networks, in which case it preserved networks with a range of performance/complexity and types trade-offs. If the network file is full and the new model is inferior to the candidate for replacement, the network set will be increased in maximum size to conform a new network.

With the help of sensitivity analysis [3], predictive error ratio was calculated for each input variable according to their degree of validity. The ratio was expressed as the relation between the predictive errors for the model with a removed given input variable and the total network errors calculated based on all input variables. This index represented that the contributions made by different input variables in predicting the outcome were ranked in order of descending importance. After the network was allowed to run, the predicted outcome was correlated with the observed outcome; and if the network predicted the outcome incorrectly, by a process of back propagation, hidden weights within the network were remodified until the predicted outcome was accurate. At last, the intelligent problem solver retained the best architecture. After training was completed, the best ANN model was validated in test set whose outcome in terms of osteoporosis was concealed. Once ANN model made a prediction for each subject, the predictive performance of the network was evaluated.

2.4. Statistical analysis

Data were expressed as mean ± standard error or ratio. To compare the differences among 3 sets, Cochran Q test and Kruskal-Wallis ANOVA test were used. The statistical significance for comparisons among 3 sets was defined as p value less than 0.05. To assess the quality of classification model in clinical study, discrimination and calibration should be calculated concurrently. Discrimination measures how well a model identifies subjects correctly as two different classes; calibration can evaluate the degree of correspondence between estimated probabilities produced by a model and actual observation. The area under the receiver operating characteristics curve (AUC) was used as a measure of a model’s discriminatory power. An AUC between 0.7 and 0.8 was classified as "acceptable" and between 0.8 and 0.9 as "excellent". The sensitivity and specificity at a cut-off value corresponding to the highest accuracy were also computed. Calibration was assessed using Hosmer-Lemeshow goodness-of-fit statistic (H-statistic) which divides subjects into deciles based on predicted probabilities and then computes a chi-square from observed and expected frequencies. A statistically good fit is defined as p value more than 0.05.

3. Results

The characteristics of study participants are listed in Table 1. Their mean age was 63.50 ± 0.24 years ranged from 50 to 91 years old and male to female ratio was 0.13. The final best ANN was a multilayer perceptron network with one input layer of 7 neurons, one hidden layer of 13 neurons, and one output layer with 1 neuron. Three variables (current

(4)

smoking, alcohol drinking, and steroid use) were removed; and remaining 7 input variables (gender, age, weight, height, BMI, postmenopausal status, and coffee consumption) were taken as significant features. Table 2 shows their ratios and ranks of sensitivity analysis for training, selection, and test sets. In quantitative determination, we averaged the ranks of three subsets for each variable as a new score Rankavg. In order of descending importance by Rankavg, 7 variables were gender, postmenopausal status, weight, height, BMI, age, and coffee consumption.

The predictive performances of ANN for 3 sets are listed in Table 3. Training and test sets had excellent discrimination with good sensitivity and specificity accordantly. On the other hand, all sets had statistically good fit represented by statistically insignificant H-statistic (p > 0.50). These results suggested that final best ANN concurrently had excellent discriminatory power and good-fit calibration.

Table 1. Comparison of training, selection, and validation sets. Training Selection Test p

Gender 90.45 88.29 86.00 < 0.001 Age 63.95 ± 0.34 62.40 ± 0.48 63.71 ± 0.50 0.02 Weight 57.47 ± 0.36 57.55 ± 0.53 58.32 ± 0.50 0.21 Height 153.58 ± 0.25 154.16 ± 0.39 154.33 ± 0.37 0.23 BMI 24.34 ± 0.14 24.19 ± 0.20 24.48 ± 0.20 0.51 Postmenopause 83.78 80.86 82.00 0.15 Steroid 8.96 6.00 6.00 0.048 Smoking 2.42 3.43 4.00 0.28 Alcohol 2.28 2.57 2.00 0.88 Coffee 4.98 6.57 4.00 0.27 BMD 1.01 ± 0.01 1.01 ± 0.01 1.01 ± 0.01 0.84 T score -1.09 ± 0.06 -1.08 ± 0.09 -1.05 ± 0.08 0.88 Osteoporosis 20.63 18.86 19.71 0.13 Gender, female ratio (%). Postmenopause, steroid, smoking, alcohol, coffee, and osteoporosis are expressed as ratio (%). The unit for age, weight, height, BMI, and BMD are years, kg, cm, kg/m2, and g/cm2, respectively.

4. Discussion

The evolution of ANN has become more satisfactory in medical research than before. A speculative benefit of this artificial intelligence is its power to identify complex correlative interactions among clinical data and final diagnosis. It can prune excessive variables during training stage and explore latent types of clinical data in biological nature. Good results by using limited variables and ANN to assess the predictability of osteoporosis in the elderly are presented in this study. For any predictive model to be practical in making clinical decision, it should use data that are easily available to clinicians at the time of triage. In order to avoid model complexity for clinicians, it may not be necessary to add more variables even though they perhaps have some influences on predictive power. In a similar study, Rae et al. used multiversion system of ANN with 20 risk factors to predict osteoporosis in 274 UK women against the measurement of quantitative ultrasonography for right heel bone [4]. Their model resulted in AUC of

(5)

0.80, but our model had simpler algorithm and structure with only 7 predictors to obtain a comparable discriminatory power.

Table 2. Sensitivity analysis among training, selection, and test subsets of best ANN. Training Selection Test

Ratio Rank Ratio Rank Ratio Rank Rankavg

Gender 1.116007 1 1.097456 1 1.222616 1 1.00 Age 1.048779 5 0.984566 5 0.993698 6 5.33 Weight 1.080328 2 0.964511 7 1.027591 3 4.00 Height 1.049166 4 1.003382 4 1.015637 4 4.00 BMI 1.047970 6 0.969516 6 1.043441 2 4.67 Postmenopause 1.052349 3 1.027129 2 1.014849 5 3.33 Coffee 1.024611 7 1.020078 3 0.959405 7 5.67

Table 3. Results of discrimination and calibration by best ANN. Training Selection Test

AUC 0.81 ± 0.02 0.75 ± 0.04 0.82 ± 0.03 p1 < 0.001 < 0.001 < 0.001 Sensitivity (%) 76.6 72.7 78.3 Specificity (%) 71.1 70.4 73.3 H-statistic 8.95 8.73 8.00 p2 0.35 0.27 0.24

p1, p value derived from AUC; p2, p value derived from H-statistic.

In many ANN investigations, data structure of input variables is seldom discussed. We can identify certain differences among three sets in Table 1. ANN was constructed from training set and internally validated successfully which represented internal generalization was well-behaved. On the other hand, ANN can extract particular input variables to assemble the model. Although variables operated in ANN model should not be explained as independent predictors as discerned by conventional statistics, they could be interpreted as part of the global function in ANN, expressing the complex nature among clinical variables. Considering 4 input variables (steroid use, current smoking, alcohol drinking, and coffee consumption) in our study might encode similar information. A model could depend completely on one, totally on the other, or on some combination of them. Then sensitivity analysis presents an associated sensibility to them. If one input variable is eliminated, the model may perform properly because other input variables still supply important information. The removed variable may be classified as of low sensitivity even though they might encrypt essential information. Therefore, our ANN selected only coffee consumption on behalf of other 3 variables. In addition, these clinical parameters are well-known having complex interrelationships that is why ANN “thought” coffee consumption could be an index variable. Moreover, a variable that encodes relatively insignificant information, but is the only variable to do so, may have higher sensitivity than any number of variables that mutually encrypt more influential information such as variable age.

The investigations of ANN to predict osteoporosis are limited in clinical medicine. In one pilot study, Ongphiphadhanakul et al. constructed an ANN to classify 129 Thai postmenopausal women into osteopenia or non-osteopenia subjects [5]. Their results were acceptable with reasonable sensitivity (76.2 ~ 80.0%) and low specificity (12.5 ~

(6)

33.3%). Although their goal and population was different with our study, the specificity was far less than ours. The size of participants and the predictors they selected were the possible reasons. Sadatsafavi et al. developed several ANN models from the dataset of 2158 Iranian postmenopausal women whose BMD values were all measured by DEXA [6]. They found that 3 to 5 input variables (age, weight, years since menopause, steroid use, estrogen use) could get good discrimination (AUC = 0.749 ± 0.017 ~ 0.818 ± 0.019) with 86.3% sensitivity and 72.1% specificity. Even though the findings and performance of their ANN model was comparable with ours, the most different emphasis was that we included the elderly whether male or female. Of note, they employed a traditional stepwise regression analysis as pre-processing method in order to select significant input variables. We believe that this superfluous procedure might lose the natural flexibility of ANN; hence, we utilized the sensitivity analysis to stress the presentation of possible results and ways to create the result in a process which involves uncertain factors by assigning different values for these factors. Through these comparisons, we can confirm that our ANN not only had appropriate design with adequate predictors but also had good performance to successfully predict osteoporosis in the elderly.

There are two limitations in this study. First, the number of female had larger ratio in our study population; this uneven distribution could lead to possible gender bias. Theoretically, ANN has its pliability to overcome this heterogeneity since no assumption of variable distribution is necessary in ANN modeling. However, osteoporosis in men is often neglected, despite the fact that one-third of hip fractures occur in men. We still need more male participants in the future study. Second, our study was carried out at a single institution. The generalization is foreseen to decline when applying any predictive tool to different populations as a result of possible confounders. Further studies in different centers for external validation should be designed to corroborate our findings and decrease potential interinstitutional variations.

5. Conclusion

Our results show that ANN could have a good performance in identifying osteoporotic subjects in the elderly and it might serve as an alternative tool to screen the individual who should arrange further work-up like DEXA. This kind of ANN approach could help clinicians to initiate other preventive actions in the future.

References

[1] Wilkins CH, Birge SJ. Prevention of osteoporotic fractures in the elderly. Am J Med 2005; 118: 1190-5. [2] Chiu JS, Chong CF, Lin YF, Wu CC, Wang YF, Li YC. Applying an artificial neural network to predict

total body water in hemodialysis patients. Am J Nephrol 2005; 25: 507-13.

[3] Hunter A, Kennedy L, Henry J, Ferguson I. Application of neural networks and sensitivity analysis to improved prediction of trauma survival. Comput Methods Programs Biomed 2000; 62: 11-9.

[4] Rae SA, Wang WJ, Partridge D. Artificial neural networks: a potential role in osteoporosis. J R Soc Med 1999; 92: 119-22.

[5] Ongphiphadhanakul B, Rajatanavin R, Chailurkit L, Piaseu N, Teerarungsikul K, Sirisriro R, Komindr S, Pauvilai G. Prediction of low bone mineral density in postmenopausal women by artificial neural network model compared to logistic regression model. J Med Assoc Thai 1997; 80: 508-15.

[6] Sadatsafavi M, Moayyeri A, Soltani A, Larijani B, Nouraie M, Akhondzadeh S. Artificial neural networks in prediction of bone density among post-menopausal women. J Endocrinol Invest 2005; 28: 425-31.

Referanslar

Benzer Belgeler

22 Şubat 2003, Cumartesi 17:00 SADBERK HANIM MÜZESİ Piyasa Caddesi 25-29, Büyükdere. &lt;s^§&gt; Vehbi

Plus loin, sur la colline, cet arôme accueillant s’accentue et vous conduit direc­ tement au charmant petit magasin de café du Kurukahveci (vendeur de café en grain)

Derenin sonunda (Dört Kardeşler) denilen dört tane asırlık ağaçların bulunduğu yere gitmedim, gitmeğe lü­ zum görmedim.. Derenin sağ tarafında yalılar,

Sonuç olarak, sürgün uzunluğu, çap kalınlığı ile aşı tutum oranlarına göre, Yer Elması tipi ile M9 ve MM106 anaçları bazı ara anaçlık özellikleri

Şuan evcil hayvan sahibi olan ve olmayan bireylerin Lexington Evcil Hayvanlara Bağlanma Ölçeği’nden aldıkları toplam ve alt boyut puanları arasında

Beklenen politika faizindeki negatif bir Ģoka hisse senetleri fiyatları ise durgunluk ve geniĢleme döneminde beklenildiği gibi pozitif tepki vermiĢ ancak

Yüksek radyasyonlu bölgede yetişen fasulyelerde, ağır metalleri bağlayarak bitkileri koruduğu bilinen sistin sintaz proteininin normal bitkilere kıyasla üç kat daha