SEARCH

SEARCH BY CITATION

Keywords:

  • ANN;
  • optimum fragmentation;
  • network system;
  • shocks;
  • stone(s)

Abstract

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. PATIENTS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. REFERENCES

OBJECTIVE

To assess the ability of artificial neural networks (ANNs) to predict optimum renal stone fragmentation in patients being managed by extracorporeal shock wave lithotripsy (ESWL).

PATIENTS AND METHODS

The study included 82 patients with renal stones who were being treated by ESWL. Data (input and output values) from 60 patients in whom there was optimum fragmentation of stones by ESWL were used for training the ANN. These data mainly included the 24-h urinary variables, the radiological features of the stone disease and the ESWL settings used. The predictability of the trained ANN was tested on 22 subsequent patients, by supplying the input variables of the 22 patients into the trained ANN and recording the output values (predicted values). After subjecting these patients to ESWL, the actual results (observed values) were recorded. The predicted and the observed values were then compared.

RESULTS

In the 22 patients in whom predictability was tested, the trained ANN predicted optimum fragmentation at ≤ 13 000 shocks/stone (as per study protocol) in 17 and optimum fragmentation at> 13 000 shocks/stone in the other five. In the 17 patients (test set) where the trained ANN had predicted optimum fragmentation at ≤ 13 000 shocks/stone, the optimum fragmentation was at that value, although the predicted and observed values were not identical. The overall correlation between the predicted and the observed values was 75.5% (correlation coefficient 0.7547) in these 17 patients. Of the other five patients, none had optimum fragmentation at < 13 000 shocks/stone, as predicted by the trained ANN, giving complete accuracy for this factor.

CONCLUSION

This was a pilot study, i.e. an initial attempt to use an ANN in this regard, and although there were few patients, such that it is not possible to make final recommendations, the overall predictability was ≈ 75%. An encouraging outcome of the study was that the trained ANN identified patients unlikely to benefit from ESWL. Using a larger dataset and identifying more significant variables, while eliminating inputs with a negative effect, the efficiency and utility of this ANN can probably be enhanced and in future it might be possible to predict stone fragmentation with reasonable accuracy.


INTRODUCTION

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. PATIENTS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. REFERENCES

Despite advances in diagnosis and therapy, renal stone disease remains a significant health problem. Two decades ago open surgery was the main treatment for symptomatic renal stones, but it is increasingly being replaced by less invasive therapies. At present ESWL, percutaneous nephrolithotomy and other endoscopic methods of stone retrieval are being used to treat> 90% of patients with renal stones. Open surgical procedures are now infrequent for treating renal stones and thus the incidence of morbidity associated with stone surgery has also markedly decreased. Since its introduction in 1980 [1] ESWL has considerably changed the management of renal stones and has become the therapeutic procedure of choice in most cases.

Artificial neural networks (ANNs) provide an ‘intelligent’ means of predicting useful outcomes with greater accuracy and efficiency. This algorithm is based on the idealised model of a biological neurone (unit) and shows great promise in overcoming the complexities in behaviour of biosystems/materials which are otherwise difficult to comprehend. Thus ANNs can be viewed as a simulation of human brain function, where sets of data in the form of input-output patterns are processed to train the ANN. The similarity between ANNs and the biological neural network lies in the parallel and distributive mode of processing the data. ANNs are created with the data from known results; the system is trained in the neural network model of a software program. Various such programs are available, e.g. MATLAB is a technical computing environment for high-performance numeric computations and visualization, and has several ‘tool-boxes’, e.g. processing control, system identification, optimization, neural networks, etc. Using a particular tool-box, different training and analysis equations can be used. For example, a feed-forward back-propagation ANN system may be used, as shown in Fig. 1; this consists of an input layer, a hidden layer and an output layer. Each layer consists of neurones or processing units, which form a medium for the forward pass of the inputs and back-propagation of the error. Each unit of the input layer is connected to each unit of the hidden layer, each of which is in turn connected to each unit of the output layer. This gives many combinations having different permutations. The input neurones, which process the data, are connected with the hidden neurones and they then, after mathematically manipulating the data, provide the predictions via the output neurones [2].

image

Figure 1. The single-layer feed-forward back-propagation ANN model.

Download figure to PowerPoint

Each variable is represented by a random numerical value. Using these values a random set of weights is initially assigned to the connections between the layers. The first output of the network is determined using these weights. The output thus obtained is compared with the actual output of the pattern pair and the mean square error calculated. An error optimization algorithm then minimizes this error. A feed-forward back-propagation ANN system has the property of self-optimization of the error during training. Thus, the final weight of a particular variable is decided by the system itself, determined precisely by the relative impact of the variable in the dataset in relation to the actual output variable. Accordingly the system has the ability to identify variables having a significant positive influence during the training and variables having minimal or negative influence during training. The negative and the least influencing variables can then be eliminated from the final training to optimize the system. The trained network is then used to provide predictions of outcome in situations where the outcome is unknown. Basically, the training of the ANN is a process of arriving at an optimum ‘weight space’ of the network. The construction and training of the ANN system needs patience and dedication to assess the influence of each variable individually, to achieve an effectively trained system. Once this is done the subsequent exercise of prediction is a relatively easy task.

Thus in the present study we assessed the ability of an ANN to predict optimum renal stone fragmentation in patients being treated by ESWL.

PATIENTS AND METHODS

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. PATIENTS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. REFERENCES

The study comprised patients with renal stones who presented to our urology department between June 2000 and December 2001, and who were managed by ESWL. Only patients with renal stones (stones in the pelvicalyceal system) were included and one stone (selected randomly) from each patient was considered; in all, 82 patients treated by ESWL were included.

Following the usual guidelines of indications and contraindications for ESWL, the patients were treated using an electrohydraulic lithotripter. As per the protocol followed at our institution, ≤ 13 000 shocks/stone, involving not more than four ESWL sittings (with a power range of 14–18 kV and a shock frequency of 60–90/min) were used to achieve optimum fragmentation, defined as no fragments (after fragmentation) of> 4 mm.

Data from 60 patients (the training set) who had had optimum fragmentation of their stones were used for training the ANN. Initially many relevant variables were assessed during training of the ANN, comprising clinical and serum variables, routine and microscopic examination of urine, 24-h urinary variables, radiological details of the stone disease (findings of IVU taken as conclusive) and the ESWL data (Table 1). After assessing the influence and relative significance of each variable during ANN training, those which had a reasonably significant influence during training were included and those with little influence excluded from the final training of the ANN. Stone size was the most influential variable, followed by the total number of shocks used and the 24-h urinary volume. Factors like gender, flank pain, routine urinary variables and the total number of stones had no significant influence during training and were therefore excluded to optimize the system.

Table 1.  The variables used for training the ANN
TypeDetail
ClinicalAge
Sex
Body habitus
± Flank pain
Laboratory
SerumCalcium
Phosphorus
Uric acid
Creatinine
Urine (routine)pH
Specific gravity
Proteins
Sugars
MicroscopicPus cells
Red blood cell
Casts
Crystals
24-hour urineVolume
Calcium
Phosphorus
Uric Acid
Oxalate
Citrate
Sodium
Magnesium
Radiological
UltrasonographyTotal number of stones
Size of particular stone
Site of particular stone
±  Hydronephrosis
IVUTotal number of stones
Size of particular stone
Site of particular stone
Radio-opacity (very  opaque, opaque,  moderately opaque)
± Hydronephrosis
ESWLTotal number of sittings
Total number of shocks
Mean power
Mean Frequency

The network was trained using the MATLAB software system; the working code for the ANN was constructed so that it was compatible with the analysis and processing of the input data. The ANN was trained using a single-layer feed-forward back-propagation network. To assess the training status of the ANN, 10 randomly selected data from the training set were used for validation (validation set), and once validation was satisfactory further training was stopped. After the ANN was considered to be reasonably trained, input variables (similar to those used for training) from subsequent patients were serially fed into the trained ANN and the number of shocks required for optimum fragmentation of each stone, as predicted by the ANN via the output data, recorded. Using the usual protocol these patients subsequently underwent EWSL and the results (observed values) were recorded. The predicted and the observed values were then compared.

During the study period, 22 patients were included to assess the predictability of the trained ANN; of these, the ANNs system predicted optimum fragmentation (≤ 13 000 shocks/stone) in 17 while in the other five the ANN predicted optimum fragmentation at> 13 000 shocks/stone. In these 17 patients the predicted and observed values were recorded for comparison, and this group thus formed the test set, while the other five patients were excluded from the test set as> 13 000 shocks/stone could not have been used. Consequently, the final observed values for comparison were not available for these five patients. In the test set of 17 patients the predicted and observed values were compared.

RESULTS

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. PATIENTS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. REFERENCES

In the 22 patients in whom the predictability of the ANN was assessed, the trained system predicted optimum fragmentation up to the maximum permissible (as per study protocol) number of shocks/stone in 17 and optimum fragmentation exceeding this in the other five. In the 17 patients (test set) where the ANN predicted optimum fragmentation at ≤ 13 000 shocks/stone this was indeed the case, although the predicted and observed values were not identical (Fig. 2). In the other five patients none had optimum fragmentation at < 13 000 shocks/stone, as predicted by the ANN.

image

Figure 2. The relationship between predicted and observed data (in the test set), with error estimates. Observed data, red closed squares; predicted data, green open circles; error, light green closed circles.

Download figure to PowerPoint

In the test set the difference in predicted and observed values was < 10% in 10 patients and < 15% in another three; in the remaining four the difference was 20–30%. The correlation between the predicted and observed values gave a correlation coefficient of 0.7547 (Fig. 3).

image

Figure 3. The regression analysis of observed and predicted values. R2 = 0.5696.

Download figure to PowerPoint

The ANN system correctly identified all five patients in whom optimum fragmentation did not occur at < 13 000 shocks/stone, giving complete accuracy (Table 2). This was one of the major achievements of the ANN prediction system.

Table 2.  ANN prediction of the actual shocks used and alternate treatment in five patients
No.ANN predicted no. of shocksActual shocksFragmentationAlternative treatment
115 91612 000PartialPCNL
215 455  9 000NonePCNL
316 27412 500PartialOpen surgery
421 00612 000NonePCNL
518 338  9 000NoneOpen surgery

DISCUSSION

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. PATIENTS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. REFERENCES

Renal stone disease is a common problem in daily urological practice and ESWL is now considered the treatment of choice for most of these cases. ESWL represents a revolutionary advance in treating patients with renal stones, but shock waves do not pass through the body and affect only the stones. There is now evidence that there is some degree of trauma to renal parenchyma with virtually every ESWL treatment [3]. Several other factors must be considered; ESWL may not be a cost-effective option compared with other available treatments. Several visits to the hospital and the medications needed during each session of ESWL means more cost must be borne by the patient. Also some form of anaesthesia or analgesia is required in most cases, especially in younger patients. As several ESWL machines involve the use of fluoroscopy for locating the stone and for focusing, some radiation exposure is inevitable [4]. Finally, in several cases the optimum fragmentation of the stone is not possible even after three or four sessions of ESWL, and an alternative treatment is required. All these factors assume particular importance when the final goal of the treatment is not achieved.

All these factors emphasize that appropriate patient selection is essential; most centres follow the usual guidelines on indications and contraindication for ESWL, mainly including stone size and the presence or absence of distal obstruction. Other factors influencing stone fragmentation are invariably not considered or may be difficult to assess. Thus, there appears to be a need for a method or system which could help to identify patients who would be best managed by ESWL and those who should be offered alternative treatments at the beginning of the treatment plan. The use of shock waves in a patient who is unlikely to benefit from ESWL could be avoided and thus wastage and any acute complications (e.g. haematuria, skin abrasions, drug hypersensitivity, and the risk of long-term complications) avoided. The need for repeated anaesthesia/analgesia and several visits to the hospital would also be eliminated.

For pattern recognition, ANNs are being used effectively in medical diagnosis. In urology ANNs have recently been applied to several diagnostic and prognostic problems, and some applications have already been offered to the urologist. Most of these studies have been conducted on prostatic disorders, especially carcinoma, and have shown acceptable results [5,6].

The present study was prospective and not randomized, assessing the role of ANNs in patients with renal stones managed by ESWL. Relatively simple variables were used for training the ANN, so that after completing the training and once the ANN was in use, similar variables would be needed to provide the required information (prediction). However, some factors seem to be significant in determining the fragmentation properties of a given stone. One such is the hardness of the stone, determined by its chemical composition. If such information is used in training the ANN better predictability would be expected. Recently there have been reports on the role of conventional and spiral CT in determining the nature of stones, based on the CT attenuation (Hounsfield) units [7]. Another preliminary study by Joseph et al.[8] involved the use of CT attenuation values for various types of renal stones to directly predict their successful fragmentation by ESWL; these results were encouraging. Therefore, it is evident that if such variables are included in the input data, a trained ANN with high predictability can be constructed.

The correlation between the predicted and observed values for the number of shocks required for optimum fragmentation (0.7547) has not been assessed previously. Michaels et al.[9], using an ANN to predict stone growth after ESWL, reported 91% accuracy for predicting stone growth. Cummings et al.[10], attempting to predict the spontaneous passage of a ureteric calculus using an ANN, reported a predictive value of 76% and a sensitivity of 100%. Volmer et al.[11] used an ANN and infra-red spectroscopy to detect the most frequent composition of urinary calculi, concluding that the ANN was more accurate than library searching and required less expert knowledge.

The slightly lower overall accuracy of the present ANN was possibly a result of the other neural inputs used in the analysis which had a negative and confounding effect on the final prediction. This could be addressed by either eliminating each neural input from the analysis or by using different combination of inputs and then re-training/re-validating the dataset for the final outcome, thus optimizing the ANN.

Currently, ESWL forms an integral part of any treatment protocol for managing renal stones, but not all patients with renal stone disease benefit from ESWL. Despite several advances in ESWL technology, the ideal lithotripter, giving maximum stone fragmentation with minimal adverse effects and no need for anaesthesia/analgesia, has yet to be developed. The present study was a preliminary exercise and although there were few patients, and thus no clear recommendations, the overall predictability was ≈ 75%. That a trained ANN was able to identify patients unlikely to benefit from ESWL was encouraging. The use of a larger dataset, and identifying more significant variables while eliminating those inputs with a negative effect, could improve the efficiency and utility of this ANN and its ability to predict stone fragmentation with reasonable accuracy.

REFERENCES

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. PATIENTS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. REFERENCES
  • 1
    Chaussy C, Brendel W, Schmiedt E. Extracorporeally induced destruction of kidney stones by means of shock waves. Lancet 1980; 2: 126570
  • 2
    Wei JT, Zhang Z, Barnhill SD, Madyastha KR, Zhang H, Oesterling JE. Understanding artificial neural networks and exploring their potential applications for the practicing urologist. Urology 1998; 52: 16171
  • 3
    Rubin JI, Arger PH, Pollack HM et al. Kidney changes after extracorporeal shock wave lithotripsy: CT evaluation. Radiology 1987; 162: 214
  • 4
    Bush WH, Jones D, Gibbons RP. Radiation dose to patient and personnel during extracorporeal shock wave lithotripsy. J Urol 1987; 138: 7169
  • 5
    Naguib RN, Robinson MC, Neal DE, Hamdy FC. Neural network analysis of combined conventional and experimental prognostic markers in prostate cancer: a pilot study. Br J Cancer 1998; 78: 24650
  • 6
    Mattfeldt T, Kestler HA, Hautmann R, Gottfried HW. Prediction of prostatic cancer progression after radical prostatectomy using artificial neural networks: a feasibility study. BJU Int 1999; 84: 31623
  • 7
    Nakada SY, Hoff DG, Attai S, Heisey D, Blankenbaker D, Poznaik M. Determination of stone composition by nonconstrast spiral computed tomography in the clinical setting. Urology 2000; 55: 8169
  • 8
    Joseph P, Mandal AK, Singh SK, Mandal P, Sankhwar SN, Sharma SK. Computerized tomography attenuation value of renal calculus: Can it predict successful fragmentation of the calculus by extracorporeal shock wave lithotripsy? A preliminary study. J Urol 2002; 167: 196871
  • 9
    Michael 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: 3358
  • 10
    Cummings JM, Boullier JA, Izenber SD, Kitchens DM, Kothandapani RV. Prediction of spontaneous ureteral calculous passage by an artificial neural network. J Urol 2000; 164: 3268
  • 11
    Volmer M, Wolthers BG, Metting HJ, Deltaan TH, Coenegracht PM, Van der slik W. Artificial neural network prediction of urinary calculus composition analyzed with infra-red spectroscopy. Clin Chem 1994; 40: 16927
Abbreviations
ANN

artificial neural network.