Predicting hypogonadism in men based upon age, presence of erectile dysfunction, and depression

Article metrics

Abstract

Hypogonadism, a disorder associated with aging, can cause significant morbidity. As clinical manifestations of hypogonadism can be subtle, the challenge and the burden of diagnosis remain the responsibility of the clinician. Four different analytic methods were used to predict hypogonadism in men based upon age, the presence of erectile dysfunction (ED) and depression. 218 men were classified by age, serum testosterone level, the presence of ED and depression. Depression was determined by the Center for Epidemiologic Studies Depression Scale (CES-D). ED was assessed by the Sexual Health Inventory for Men (SHIM). Hypogonadism was defined as a serum testosterone level <300 ng/dl. An artificial neural network (ANN) was programmed and trained to predict hypogonadism based upon age, SHIM, and CES-D scores. Subject data was randomly partitioned into a training set of 148 (67.9%) and a test set of 70 (32.1%). The ANN processed the test set only after the training was complete. The discrete predicted binary output was set to (0) if testosterone level was <300 ng/dl or (1) if >300 ng/dl. The data was also analyzed by standard logistic regression (LR), linear and quadratic discriminant function analysis (LDFA and QDFA, respectively). Reverse regression (RR) analysis evaluated the statistical significance of each risk factor. The ANN can accurately predict hypogonadism in men based upon age, the presence of ED, and depression (receiver-operating characteristic=0.725). A four hidden node network was found to have the highest accuracy. RR revealed the depression index score to be most significant variable (P=0.0019), followed by SHIM score (P=0.00602), and then by age (P=0.015). Hypogonadism can be predicated by an ANN using the input factors of age, ED, and depression. This model can help clinicians assess the need for endocrinologic evaluation in men.

Introduction

Androgen deficiency can occur at any time during a man's life, but occurs more frequently with advancing age: it is estimated to affect between 19–34% of men over the age of 60.1 Over time, lack of testosterone has the potential to cause significant health problems. Clinical manifestations may include decreased libido, erectile dysfunction (ED), musculoskeletal decline, fatigue, irritability, cognitive impairment, and even depression. In older males, the scenario of androgen deficiency has been referred to as andropause, late-onset hypogonadism, or androgen deficiency in the aging male (ADAM).2 Androgen deficiency can be challenging for the clinician to identify as a patient may manifest only one of the signs of the disorder. As life expectancy increases, more men are likely to seek medical attention for a myriad of medical issues. At either the primary care or subspecialty level, the opportunity will therefore exist for the physician to encounter a male with hypogonadism. Identification and treatment of this disorder has the potential to ameliorate the symptoms of depression, fatigue, or ED, as well as increase quality of life.3 As a variety of effective testosterone replacement modalities exist, the challenge and burden of diagnosis remain the responsibility of the clinician.

Our study seeks to determine if hypogonadism can be predicted by evaluating three patient characteristics that can be easily ascertained during an outpatient clinic encounter: the presence of depression, ED, and advancing age.

Three standard linear analytic methods were used to investigate the hypothesis: logistic regression (LR), linear and quadratic discriminant function analysis (LDFA and QDFA, respectively). An artificial neural network (ANN) also analyzed the data to ascertain whether the association between hypogonadism and these three risk factors existed in a nonlinear relationship. ANNs are complex computer programs that simulate the physiologic processing of a biologic neural pathway. Data is presented to a nodal system in which each node (acting as a neuron) applies a transfer function for the information it processes and presents its outcome to other nodes in a pathway. This is analogous to a neural pathway in which each neuron processes information and transfers it to one or more neurons. Nodal pathways, just as neural pathways, can exist in parallel and then can converge to yield the processed information to yet another node: allowing the information to be summarized in order to lead to a final outcome. When two or more pathways converge upon a single node, each pathway can confer different amounts of influence, positive (excitatory function) or negative (inhibitory function), on the converging node. The advantage of the ANN is that the program uses an established training data set to learn the association among several factors (in our study age, depression, and ED) and the stated output (hypogonadism). The training set has known values for the variables as well as the designated output. As it evaluates these associations, the model teaches itself to predict the outcome. It then compares each prediction to the true association and calculates the degree of error in each prediction. The degree of error is then incorporated into the program and used when analyzing new data. This allows the program to learn as it processes data, analogous to learning based upon experience. ANNs can be accessed via the internet, downloaded to PDAs or PCs and easily used by clinicians while evaluating patients.

Materials

A total of 218 men were identified in a university-affiliated clinic (average age 53.6±14.2 years; range 18–82). We used the Center for Epidemiologic Studies Depression Scale (CES-D) to asses depression.4 The Sexual Health Inventory for Men (SHIM) was used to assess ED.5 Age, serum testosterone level, SHIM, and CES-D scores were ascertained for each man. In all, 148 (67.9%) men were randomly selected to a training set and the remaining 70 (32.1%) were randomly selected to a test set, with similar outcome frequencies preserved in each set.

The SHIM survey is a validated questionnaire taken from the International Index of Erectile Function (IIEF). It is designed to screen for ED by assessing: (1) erectile rigidity, (2) maintenance of an erection after penetration/intercourse completion, (3) confidence in achieving/maintaining an erection, and (4) intercourse satisfaction. A lower value conveys more severe ED.5

The CES-D is a validated, self-administered 20-question survey that assesses depressive symptoms on a 4-point scale: the cutoff for depression is usually a score >16. The scale was developed for use in studies to ascertain the epidemiology of depressive symptoms with good internal consistency and reproducibility.4

Statistical analysis

An ANN was designed using our neural computation network for UROlogical numerical (neUROn++) models programmed in C++: a feed-forward model with one hidden layer, output unit, and a cross-entropy error function.6 The training method utilized was backpropagation.7, 8 Our program first used a pattern recognition task to ‘train’ the ANN with observed input–output pairs in a training set of patients. It then used previously unseen input–output pairs from a new patient group (test set) to ‘test’ the ‘trained’ model. The three input variables used were: (1) age, (2) CES-D score, and (3) SHIM score. The single, discrete predicted binary output was set to 0 if serum total testosterone level was <300 ng/dl or 1 if >300 ng/dl. Several architectures were investigated, including one, two, three, four, and five hidden nodes. Overlearning occurred when additional hidden nodes decreased accuracy. Evaluation of error surface properties ensured that the parameter estimates were a strict local minimum of the cross-entropy error function.9 The network was considered trained when the error was oscillating at a local error minimum. The data was also evaluated using linear analysis techniques: LR, LDFA, and QDFA. This allowed comparison of the nonlinear method of neural computation with traditional linear statistical methods. We computed receiver-operating characteristic (ROC) curve areas statistically using the statistical method described by Wickens.10

To determine the significance of each risk factor variable on the ANN, a reverse regression (RR) model was performed utilizing a Wilk's generalized likelihood ratio. This employs sequentially removing each variable from the ANN. The network is retrained without that variable input. The final error of the variable-deficient network is then compared to the error of the complete network. The probability of a null hypothesis (that the variable-deficient network and the complete network perform with the same degree of error) being rejected is calculated. This determines whether that variable is necessary for the model.9

Results

Mean testosterone was 420.7 ng/dl (±224.4). In the training group, the average age, CES-D, and SHIM scores were: 53.7 (±14.4), 16.4 (±9.5), and 10.9 (±6.4) years, respectively. In the test group, the average age, CES-D, and SHIM scores were: 53.4 (±13.7), 14.8 (±8.0), and 11.7 (±6.5) years, respectively. In the training group, 44 (29.7%) of patients had a CES-D score 22, which was considered a sign of overt clinical depression: 29 (19.6%) patients had a score within 16–21, signifying a reactive depression. In the test group, 10 (14.3%) patients had a CES-D score 22 and 18 (25.7%) patients had a score between 16–21.

A one hidden layer, four hidden node architecture was found to most accurately predict hypogonadism while maintaining an acceptable goodness of fit. The ROC curve was calculated for each analytic modality (Table 1). A value of 1.0 would reflect that the analytic modality was perfect in its classification, whereas an ROC value of 0.5 would reflect the modality was no better than chance in its predictions.9 The ROC curve for our ANN was 0.725 in both the training set and test set. LR, LDFA, and QDFA had unacceptable ROCs for both the training and test sets (Table 1). ROCs were compared using the method described by DeLong (Table 2).11 RR revealed that all three variables, when independently removed, significantly deteriorated the network's classification accuracy (Table 3). The presence of depression had the most significant impact (P=0.0019). The presence of ED had the second most importance (P=0.00602). Age had the least influence on the accuracy of the ANN, but was still significant (P=0.015).

Table 1 Model accuracies
Table 2 DeLong's ROC comparison
Table 3 Regression analysis

Discussion

Hypogonadism has been associated with depression, ED, and advancing age. The clinical manifestation may include fatigue, irritability, decreased cognitive function, osteoporosis, a reduction of muscle mass, or body fat redistribution. Although the incidence of hypogonadism increases with age, it may occur during any time of an adult male's life. The cause may be multifactorial: environment, alcohol, nutritional status, and chronic disease conditions can affect testosterone levels.12 The identification of hypogonadism based upon often vague and seemingly unrelated patient symptoms or signs can be challenging to a clinician. Only one of the signs or symptoms may become apparent during a patient encounter and, often, this can be attributed to etiologies other than testosterone deficiency. ED can result from underlying vascular disease or have a psychogenic component. Idiopathic depression by itself can account for decreased libido, fatigue, and changes in affect: symptoms also seen in hypogonadism. The clinician must maintain a low threshold to place hypogonadism within the differential diagnosis when one or more of these clinical manifestations surface. Depression, ED, and age are factors that can be easily screened for during an office visit. We report the use of an ANN as an effective tool in predicting hypogonadism based upon these three factors.

There is a decline in serum testosterone levels in men as they age. It is estimated that up to 34% of men over the age 60 will have low testosterone levels.1 Testosterone can circulate in a free form or be bound by albumin as well as sex-hormone-binding globulin: free testosterone and albumin-bound testosterone being most bioavailable. As one ages, the serum testosterone concentration is thought to decline by 3.2 ng/dl/year.1 The decrease can be attributed to several factors: an increase in sex hormone binding-globulin, hypothalamic–pituitary axis changes, and a decrease in the quality as well as the quantity of Leydig cells.13, 14 Our analysis revealed that age was an important risk factor when using the ANN to predict hypogonadism (P=0.015).

Decreasing testosterone levels may play a factor in the muskuloskeletal changes associated with aging. Chronic androgen deficiency can lead to a decline in muscle strength, increased visceral fat, osteoporosis, as well as an increased fall risk secondary to a decrease in static and dynamic balance.3, 15 Additionally, hypogonadism has shown to decrease the ability of elderly men to stand up from a chair or participate in tandem walking.14 The clinical impact could be the increased risk of fractures. Testosterone replacement has been shown to improve muscle strength in hypogonadal men. Additionally, replacement has been shown to increase lean body weight and decrease total as well as percent fat mass.3 Identification and treatment of hypogonadism, therefore, has the potential to ameliorate some musculoskeletal changes of aging.

The prevalence of ED has been shown to increase with age, and has been estimated to be as high as 61% in men age 70 or older.16 Although the precise mechanism remains unclear, testosterone is thought to influence libido and erectile function. Bioavailable testosterone concentration has been found to be positively correlated with erectile function.17 In addition, free testosterone concentrations have been shown to correlate with vascular compliance and trabecular smooth muscle relaxation in the penis, although causation has not been established.18 Intracavernosal testosterone levels have also been shown to be higher during the erect state when compared to the flaccid state, which suggests a testosterone–androgen receptor-binding mechanism.19 Furthermore, testosterone replacement has shown to improve sexual performance in certain hypogonadal men.3 It remains unclear, however, why some castrate men maintain some erectile ability. One study found castrate men who retained potency to have significantly higher baseline testosterone levels than completely impotent men who were also castrate, indicating a testosterone-dependent process.20 Even though the etiology of ED is often multi-factorial or can be the result of another disease process, the evaluation should include a serum testosterone level. Some advocate use of a free testosterone level when screening for ED. Our use of the total testosterone level to diagnose hypogonadism was based upon the FDA acceptance of the total testosterone level as a primary efficacy end point for recent testosterone replacement clinical trials.21 If a low testosterone level is the underlying cause for a patient's ED, androgen replacement may potentially improve the degree of erectile function.3 In our study, the presence of ED was determined by the SHIM survey: a subset of questions taken from the IIEF.5, 22 ED was a statistically significant risk factor in the evaluation of hypogonadism by the ANN (P=0.00602).

It is becoming more evident that men with androgen deficiency can also suffer from other nonspecific problems: depression, fatigue, anorexia, and cognitive impairment. Although the pathophysiology remains unclear, depression has been associated with testosterone deficiency. In a retrospective study by Shores et al., hypogonadal men over 45 years of age had a 3.5-fold risk increase in the incidence of depression when compared to eugonadal men.23 It has been found that cerebral perfusion itself maybe altered in hypogonadism. Testosterone replacement in hypogonadal men has been shown to increase blood flow to the midbrain, superior frontal gyrus, and midcingulate gyrus.24 It is unclear if these altered perfusion patterns account for the affective manifestations of androgen deficiency, especially depression. Although large, randomized, double-blind placebo-controlled trails are necessary, testosterone replacement has been found to be effective in treating depression in certain hypogonadal men.25 Additionally, testosterone replacement has been shown to improve positive mood and decrease negative mood.3 It can often be difficult to ascertain whether a patient's depression is idiopathic or the result of an underlying medical condition such as androgen deficiency. What makes the diagnosis of hypogonadism difficult is that idiopathic depression alone can account for fatigue, decreased libido, and psychogenic ED. Therefore, prudent clinical judgment is essential when ascertaining the etiology and choosing the treatment once the diagnosis of depression has been made. It remains to be determined whether hypogonadism is causative or merely associated with depression. The presence of depression, however, should increase one's level of suspicion for hypogonadism in the appropriate patient population. The presence of depression had the greatest influence on the accuracy of our ANN when predicting hypogonadism, more so than ED and age (Table 3).

The clinical presentation of hypogonadism can be varied and the diagnosis challenging. We have modeled an ANN program based upon the data collected from a pre-existing patient population in which each variable (age, CES-D and SHIM scores) and the outcome variate (hypogonadism) are known. RR analysis determined that all the three variables were necessary in this well-conditioned multivariate model: depression had the most significance on the accuracy of our model, then ED, and followed by age.

Our ANN can be used to predict hypogonadism based upon patient age, CES-D, and SHIM scores. A man with all risk factors will be predicted to have a high risk of hypogonadism. A man with milder ED or depression (based upon respective CES-D and SHIM scores) will have a lower probability than a similar-aged man with more significant ED or depression. Likewise, an older man without the other risk factors will be predicted to have a higher chance of hypogonadism than a younger man without ED and depression. Our analysis revealed that the association of these variables to hypogonadism exists in a nonlinear relationship, as standard linear analysis was inferior to the ANN (Table 1).

Conclusion

Our ANN provides the clinician with a superior tool one can use to predict hypogonadism based upon age, the presence of ED, and depression. Diagnosis and treatment can improve the quality of life of many men who are afflicted by this disorder.

References

  1. 1

    Harman SM, Metter EJ, Tobin JD, Pearson J, Blackman MR . The Baltimore Longitudinal Study of Again: Longitudinal effects of aging on serum total and free testosterone levels in health men. J Clin Endocrinol Metab 2001; 86: 724–731.

  2. 2

    Morales A . The andropause: bare facts for urologists. Br J Urol 2003; 91: 311–313.

  3. 3

    Wang C, Swedloff RS, Iranmanesh A, Dobs A, Snyder PJ, Cunningham G et al. Transdermal testosterone gel improves sexual function, mood, muscle strength, and body composition parameters in hypogonadal men. J Clin Endocrinol Metab 2000; 85: 2839–2853.

  4. 4

    Radloff LS . The CES-D Scale: a self-report depression scale for research in the general population. Appl Psychol Meas 1977; 1: 385–401.

  5. 5

    Capperelli JC, Siegal RL, Glasser DB, Osterloh IH, Rosen RC . Relationship between patient self-assessment of erectile dysfunction and the sexual health inventory for men. Clin Ther 2001; 3: 1707–1719.

  6. 6

    Fiesler E, Beale R, Axelrod T, Blayo F, Cios KH, Doerschuk PI et al. Handbook of Neural Computation. New York: IOP Publishing and Oxford University Press, 1997, pp. G5 4: 1–6.

  7. 7

    Werbos P . Beyond regression: new tools for prediction and analysis in the behavioral sciences. PhD Thesis. Cambridge, MA: Harvard University, 1974.

  8. 8

    Golden RM . Mathematical Methods for Neural Network Analysis and Design. Cambridge, MA: MIT Press, 1996.

  9. 9

    Michaels EK, Niederberger CS, Golden RM, Brown B, Cho L, Hong Y . Use of a neural network to predict stone growth after shock wave lithotripsy. Urology 1998; 51: 335–358.

  10. 10

    Wickens TD . Elementary Signal Detection Theory. New York: Oxford University Press, 2002.

  11. 11

    DeLong ER, DeLong DM, Clarke-Pearson DL . Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988; 44: 837–845.

  12. 12

    Wespes E, Schulman CC . Male andropause: myth, reality, and treatment. Int J Impot Res 2002; 14: S93–S98.

  13. 13

    Rubens R, Dhont M, Vermeulen A . Further studies on Leydig cell function in old age. J Clin Endocrinol Metab 1974; 39: 40–45.

  14. 14

    Neaves WB, Johnson L, Porter JC, Parker Jr CR, Petty CS . Leydig cell numbers, daily sperm production, and serum gonadotropin levels in aging men. J Clin Endocrinol Metab 1984; 59: 756–763.

  15. 15

    Szulc P, Claustrat B, Marchand F, Delmas PD . Increased risk of falls and increased bone resorption in elderly men with partial androgen deficiency: the MINOS study. J Clin Endocrinol Metab 2003; 88: 5240–5247.

  16. 16

    Bacon CG, Mittleman MA, Kawachi I, Giovanucci E, Glasser DB, Rimm EB . Sexual function in men older than 50 years of age: results from the health professional follow-up study. Ann Intern Med 2003; 139: 161–168.

  17. 17

    Tsujimura A, Matsumiya K, Matsuoka Y, Takahashi T, Koga M, Iwasa A et al. Bioavailable testosterone with age and erectile dysfunction. J Urol 2003; 170: 2345–2347.

  18. 18

    Aversa A, Isidori AM, De Marino MU, Caprio M, Fabbrini E, Rocchietti-March M et al. Androgens and penile erection: evidence for a direct relationship between free testosterone and cavernous vasodilation in men with erectile dysfunction. Clin Endocrinol 2000; 53: 517–522.

  19. 19

    Becker AJ, Ückert S, Stief CG, Truss MC, Machtens S, Friedemann S et al. Cavernous and systemic testosterone levels in different phases of human penile erection. Urology 2000; 56: 125–129.

  20. 20

    Greenstein A, Plymate SR, Katz PG . Visually stimulated erection in castrated men. J Urol 1995; 153: 650–652.

  21. 21

    Wald M, Seftel AD, Ross LS, Mohamed MAB, Niederberger CS . Computational models for detection of erectile dysfunction. J Urol 2005; 173: 167–174.

  22. 22

    Rosen RC, Riley A, Wagner G, Osterloh IH, Kirkpatrick J, Mishra A . The International Index of Erectile Dysfunction (IIEF): a multidimensional scale for the assessment of erectile dysfunction. Urology 1997; 49: 823–830.

  23. 23

    Shores MM, Sloan KL, Matsumoto AM, Moceri VM, Felker B, Kivlahan DR . Increased incidence of diagnosed depressive illness in hypogonadal older men. Arch Gen Psychiatry 2004; 61: 162–167.

  24. 24

    Azad N, Pitale S, Barnes WE, Friedman N . Testosterone treatment enhances regional brain perfusion in hypogonadal men. J Clin Endocrinol Metab 2003; 88: 3064–3068.

  25. 25

    Siedman SN, Rabkin JG . Testosterone replacement therapy for hypogonadal men with SSRI-refractory depression. J Affect Disord 1998; 48: 157–161.

Download references

Author information

Correspondence to A Kshirsagar.

Rights and permissions

Reprints and Permissions

About this article

Cite this article

Kshirsagar, A., Seftel, A., Ross, L. et al. Predicting hypogonadism in men based upon age, presence of erectile dysfunction, and depression. Int J Impot Res 18, 47–51 (2006) doi:10.1038/sj.ijir.3901369

Download citation

Keywords

  • hypogonadism
  • neural network
  • erectile dysfunction
  • depression
  • testosterone

Further reading