The use of Data Mining in the categorization of patients with Azoospermia

INTRODUCTION

Approximately 15% of couples are unable to conceive after one year of unprotected intercourse. A male factor is solely responsible in about 20% of infertile couples and contributory in another 30-40%.¹ The World Health Organisation suggests specific diagnostic algorithms for the classification and treatment of male infertility.² Men with azoospermia constitute a special group of male infertility. Azoospermia is defined as the complete absence of sperm from the ejaculate, even after centrifugation.

Azoospermia is a laboratory finding and not a diagnosis; therefore, it can be the result of a wide spectrum of pathophysiological conditions that call for different therapeutic approaches.

Knowledge discovery in databases is the computing process of finding useful hidden information in large datasets.^3-5 Data Mining is one step during the process of knowledge discovery. It consists of mining algorithms which reproduce a set of patterns that can be further applied to other databases. Statistics, machine learning, database techniques, pattern recognition, and optimization techniques are the main tools of Data Mining.^3-5

The aim of this study was to compare Data Mining diagnosis with clinical diagnosis by applying a Data Miner (DM) to a clinical dataset of infertile men with azoospermia.

SUBJECTS AND METHODOLOGY

Clinical notes of 3000 infertile men evaluated at the outpatient clinic of the Unit of Reproductive Endocrinology from 1990 to 2004 were studied retrospectively. Of these subjects, 204 men with azoospermia were identified (6.8%). Full history was retrieved from 147 of these men. A database consisting of eleven clinical and laboratory parameters was constructed: 1) age, 2) referral reason (presenting complaint), 3) personal andrologic history, 4) presence of secondary sex characteristics, 5) mean testicular volume, 6) presence of varicocele, 7) follicle stimulating hormone (FSH) serum levels, 8) luteinizing hormone (LH) serum levels, 9) testosterone (T) serum levels, 10) sperm volume, and 11) cytopathology findings from testicular fine needle aspiration biopsies (FNA).

The cohort of azoospermic men was clinically classified into four subgroups: a) eugonadal men with obstructive azoospermia (n=63, 42.8%), b) eugonadal men with non-obstructive azoospermia (n=71, 48.3%), c) hypergonadotropic hypogonadal men (n=2, 1.4%), and d) hypogonadotropic hypogonadal men (n=11, 7.5%) (Table 1 ).

Secondary sexual characteristics, mean testicular volume, serum levels of FSH, LH and T were clinically considered as primary factors for Data Mining, whereas the remaining parameters were considered as secondary factors for the computing process. Cytopathology findings were excluded from further analysis because of the significant number of missing data (34/147, 23%).

The data from the patients files were manually entered in a computer database (Microsoft Excel for Windows^®, Microsoft Office 2003). Finally, the database was transferred from Excel to a DataBase2 file so that it be compatible with the DM (DB2/Intelligent Miner for Data ®, IBM, version 6.1).

DM was specifically programmed to form four subgroups of men (clustering model with limited number of produced groups). By this means, comparison with the four clinical groups would be more feasible.

RESULTS

The Data Mining results are presented in Table 2 . The four-cluster model produced four groups of azoospermic men.

The first group (Group 1, n=86, 58.5% of the study population) consisted of eugonadal men with normal testicular volume (median 22.0 cm³) and normal serum levels of FSH, LH and T (8.1 IU/l, 1.5 IU/l, and 17.0 nmol/l, respectively). During the Data Mining, 59 men with obstructive azoospermia (93% of the total men clinically categorized as obstructive azoospermia), 26 men with non-obstructive azoospermia, and one man with hypogonadotropic hypogonadism were allocated in this group.

The second group (Group 2, n=29, 19.7% of the study population) consisted of eugonadal men with decreased testicular volume (median 6.5 cm³), raised FSH and LH serum levels, and normal serum T levels (26.4 IU/l, 15.3 IU/l, and 13.6 nmol/l, respectively). All men allocated in this group had non-obstructive azoospermia according to the clinical classification.

The third group (Group 3, n=20, 13.6% of the study population) consisted of eugonadal men with moderately decreased testicular volume (median 14.5 cm³), slightly raised FSH and LH serum levels, and normal serum T levels (18.6 IU/l, 12.0 IU/l and 19.2 nmol/l, respectively). Four men clinically classified as obstructive azoospermia and 16 men as non-obstructive azoospermia were allocated into this group.

Finally, the fourth group (Group 4, n=12, 8.2% of the study population) consisted of hypogonadal men with decreased testicular volume (median 7.0 cm³), normal FSH and LH serum levels, and low serum T levels (6.3 IU/l, 7.2 IU/l and 4.6 nmol/l, respectively). Ten out of the 11 clinically classified as hypogonadotropic hypogonadal men and the two hypergonadotropic hypogonadal men were allocated into this group.

Overall, DM concordance rate in hypogonadal men was 92% (clinical: 13 patients vs. DM – Group 4: 12 patients), in obstructive azoospermia 73% (clinical: 63 patients vs. DM – Group 1: 87 patients), and in non-obstructive azoospermia 69% (clinical: 71 patients vs. DM – Groups 2 and 3: 49 patients).

DISCUSSION

The overall performance of the DM on the andrology database was satisfactory from a technical point of view. The DM ran flawlessly over the database and reproduced a four-cluster model that could be compared with the clinical classification of the azoospermic men.⁶ A successful Data Mining procedure would classify the azoospermic men in a way similar to that of the clinical diagnosis. The DM successfully formed the groups of men with hypergonadotropic hypogonadism (Group 4) and obstructive azoospermia (Group 1). Regarding the men with non-obstructive azoospermia (Groups 2 and 3), there was only partial agreement of the Data Mining results with the clinical classification. Moreover, the DM's classification was generally characterised by low sensitivity as compared to the clinical classification, with the exception of Group 4.

An instance of significant success of the DM would be the exclusive inclusion of eugonadal men with non-obstructive azoospermia in Group 2. Nevertheless, less than 40% (29 of 71) of the total eugonadal men with non-obstructive azoospermia were included in this group; thus the classification was characterised by high sensitivity but low specificity.

Group 3 of the Data Mining included four eugonadal men with obstructive azoospermia and 16 eugonadal men with non-obstructive azoospermia. The inclusion of these two subgroups of men into the same DM group has no clinical relevance. The same also applies to Group 1.

Finally, Group 4 of the Data Mining included two hypergonadotropic hypogonadal men and 10 of the 11 hypogonadotropic hypogonadal men. This group of patients was fairly similar to the clinical grouping of the hypogonadal men. The error of the Data Mining was the allocation of both the hypergonadotropic and the hypogonadotropic hypogonadal men into the same group.

In spite the above-mentioned discrepancies, the ability of the Intelligent Miner for Data to reproduce clinically justifiable results is obvious. Each one of the groups that was formed by the Data Miner consisted of men with similar clinical features. A likely explanation for the differences between the clinical and DM classification is the failure of the mining model to conclusively use the secondary clinical parameters included in the database.

It is true, however, that the database that was used for the Data Mining process was rather small. Commercial Data Miners extract useful new information from datasets of thousands or millions of records.⁷ The relatively small size of the cohort that was used in this study was definitely a limitation for a reliable interpretation of the results. The overall impression, however, is that the application of the DM in a larger dataset would further improve its performance and data interpretation.

The most important step during Data Mining in databases is the analysis and the interpretation of the results. DMs produce numerous algorithms, models, correlations, prognostics, and decision trees, all of which have to be clinically assessed and validated. The clinician should initially go through this time-consuming evaluation and then choose the application that would be functional in the clinical situation to be dealt with.⁸

It seems that certain prerequisites must be fulfilled for a DM to be able to offer clinical assistance. A DM can successfully operate only in complete and accurate patient databases. The clinician could then use the DM to get help in decision support systems, research projects, or risk management issues (i.e. in the discovery of associations that could assist in the prevention of health-professional errors).^3,5,9

To date, Data Mining has been sparsely used in the field of Andrology. Dzeroski et al reported a database consisting of clinical details from infertile men that was analysed with the use of a DM.^10,11 The researchers investigated 382 infertile men by using 177 gene markers in order to detect Y-chromosome microdeletions and establish a correlation between genotype and phenotype. Clustering and decision-tree making were the main Data Mining tools used in this study. Data Mining suggested a group of 13 gene markers that were responsible for more than 90% of the infertile male phenotype. The authors, however, did not try to compare clinical with Data Mining findings.¹⁰ In the present study, the main aim was the comparison of the Data Mining findings with the clinical classification.

To the best of our knowledge, this is the first attempt to clinically validate the use of Data Mining in azoospermia. Our data suggest that, given the Data Mining tools we used in this study, the clinical and laboratory data available are already sufficient for the experienced andrologist to reach a diagnosis, and further assistance from a computer does not optimize diagnostic ability. Nevertheless, DM could have a potential role in education and in supporting decisions of trainees.

In this dataset, making a diagnosis without human involvement did not appear feasible. Given the poor performance of the DM in the learning set, any further examination on a new group of patients would most likely yield results of equally low value. Potential application of DM would be to unravel a hidden layer of data, i.e. overlooked pathophysiological connections. However, this application does not appear currently achievable, as the computer algorithm is proprietary and hidden from the user. Further trials are required to determine the clinical significance of the four computer generated groups.

In conclusion, in this study a database of azoospermic men was initially built and DM was used subsequently to extract useful clinical information. The Data Mining results were compared with the clinical classification of the patients. The comparison suggests that DM in its present form is not capable of consistently helping the clinician in this particular entity. The small sample size may be an important limitation. It is quite possible that Andrology, Assisted Reproduction Techniques (ART), gene therapy, and molecular biology are fields that could benefit from the use of DMs. It remains to be proven, however, whether these hi-tech products will in the future secure a place in everyday clinical practice.

REFERENCES
1. The Male Infertility Best Practice Policy Committee of the American Urological Association and The Practice Committee of the American Society for Reproductive Medicine, 2004 Report on evaluation of the azoospermic male. Fertil Steril 82: Suppl 1: S131-S136.
2. WHO Manual for the standardized investigation, diagnosis and management of the infertile male, 2000 Cambridge University Press.
3. Cios KJ, 2000 Medical data mining and knowledge discovery, IEEE. Eng Med Boil Mag 19: 15-16.
4. Stilou S, Bamidis PD, Maglaveras N, Pappas C, 2001 Mining association rules from clinical databases: an intelligent diagnostic process in healthcare. Medinfo 10: 1399-1403.
5. van Bemmel JH, Musen MA 1997 Handbook of Medical Informatics. In: Springer (ed) pp, 362.
6. Papadimas J, Papadopoulou F, Ioannidis S, et al, 1996 Azoospermia: clinical, hormonal, and biochemical investigation. Arch Androl 37: 97-102.
7. DB2 Intelligent Miner for Data, IBM. Web site: http://www-306.ibm.com/software/data/iminer/. Accessed on 27/10/2005.
8. Babic A, 1999 Knowledge discovery for advanced clinical data management and analysis. Stud Health Technol Inform 68: 409-413.
9. Greenes RA, 2002 Future of medical knowledge management and decision support. Stud Health Technol Inform 80: 29-44.
10. Dzeroski S, Hristovski D, Kunej T, Peterlin B, 2000 A data mining approach to the development of a diagnostic test for male infertility. Stud Health Technol Inform 77: 779-783.
11. Dzeroski S, Hristovski D, Peterlin B, 2000 Using data mining and OLAP to discover patterns in a database of patients with Y-chromosome deletions. Proc AMIA Symp pp, 215-219.

Address correspondence and requests for reprints to:
Dr. Dimitrios G. Goulis, 1^st Department of Obstetrics
& Gynaecology, "Papageorgiou" General Hospital,
Periferiaki Odos, Nea Efkarpia, 564 03, Thessaloniki, Greece,
Tel.: +30 2310 693384, Fax: +30 2310 991510,
e-mail: dgg30@otenet.gr

Received 21-09-05, Revised 15-10-05, Accepted 20-10-05