P-057 Machine learning-based prediction of testicular sperm extraction: comparison of different preprocessing and models, required sample size and relevance of input biomarkers

Study question

Can advanced machine learning applied to the preoperative assessment predict the testicular sperm extraction outcome in azoospermic context and how many patients are required?

Summary answer

Despite encouraging results (AUC = 92.0%, sensitivity = 83.9% and specificity = 84.2%), integrating new biomarkers would probably be more relevant than enrolling additional patients.

What is known already

Testicular sperm extraction (TESE) is an essential therapeutic tool for the male infertility management and is often the “last hope” before gamete donation for these patients. However, it is an invasive procedure and is successful in up to 50%. Until now, no model is sufficiently powerful to accurately predict the success of sperm retrieval in TESE. Among the few models already developed, the findings are highly disparate despite having common input data (preoperative assessment). Moreover, only few types of machine learning models and procedures have been investigated. Performances were mostly capped despite the inclusion sometimes of more than 1000 patients.

Study design, size, duration

Data of 175 patients who underwent TESE between 2012 and 2021 were retrospectively analyzed. The performances of a wide range of preprocessing methods and machine learning models (state-of-the-art methods in machine learning) we explored, evaluated, and compared. The objective was to predict the presence or absence of spermatozoa, using 17 parameters (clinical, hormonal, genetic, history) from the preoperative assessment. The study protocol was approved by a local ethics committee (IRB CER-2021-041).

Participants/materials, setting, methods

After data preprocessing (standardization…), Machine Learning models (Bayesian Naive Classification, logistic regression, k-nearest neighbor classifier, support vector machine, random forests, GradientBoosting and XGBoost) and Deep Learning models were tested. The validation procedure consisted of splitting the dataset into a training set and test set. Beyond the standard metrics (sensitivity, specificity, AUC-ROC), the identification of the most relevant variables and the learning curve to determine the optimal patient number to be included were performed.

Main results and the role of chance

At least one live spermatozoon was found in the testicular tissue of 104 (59.4%) patients (positive TESE) out of 175. The best performing model (Random Forest with appropriate preprocessing) obtained the following results on the test set: AUC = 92.0%, sensitivity = 83.9% and specificity = 84.2%, leading to an efficient tool, which gives additional and more relevant information than the different variables taken separately. Inhibin B, FSH and history of cryptorchidism were the variables with the most discriminating power. However, a plateau in the model performance was observed (beyond 110 patients), whatever the approach or the preprocessing used. A trend curve shows that beyond 110 patients, no improvement can be observed and cast doubt about the power of the traditional preoperative parameters assessed before TESE. The classic preoperative assessment can probably not fully predict the TESE outcomes. Further work is needed to be enhance with new hypothesis and the use of new biomarkers to be integrated into the models.

Limitations, reasons for caution

The main limitation was the monocentric design and the use of retrospective data.

Wider implications of the findings

Machine learning models can provide the basis for an enhanced decision support system tool in the context of azoospermia. Indefinitely increasing the number of participants is not likely to be the solution: further hypotheses and biomarkers integration into the models will probably be necessary to improve performance.

Trial registration number

not applicable

[Asynchronous sequential stimulation. A new signal treatment for cochlear implants]

Cochlear implants use a fixed or FO dependent stimulation rate, whatever the sound analysed (vowel or consonant) and its analysis procedure (fixed filters or FFT). We present a procedure which varies as a function of the nature of the input signal. This method is based on FFT analysis using a variable with analysis window. Short windows are applied on the transient part of the signal, providing poor frequency resolution but good time resolution; they accurately permit to follow the speech signal in time during its fast temporal variations. Large windows are applied on the stationary parts of the signal, providing better frequency resolution, but poor time resolution. Transient parts of the speech order this window switching; they are detected using the statistical properties of the FFT; moreover the narrow windowing is coupled with an increase of the stimulation frequency. This strategy has been implemented using the Digisonic cochlear implant software, and clinically assessed on 6 regular cochlear implant users, owing to a consonant-vowel-consonant test. This Asynchronous-Interleaved-Stimulation (AIS) strategy provides the patients with better discrimination than fixed window FFT analysis. The coding protocol are described and results presented.

[Digital multi-filter auditory prosthesis: study of its efficiency in function of number of filters and their programming]

We designed a prototype of a 7-filter digital auditory prosthesis table prototype. For each of these filters frequency band width, amplification and compression were programmable in order to adapt these parameters to the deaf patient's audiometric particularities. We compared the hearing improvement it was possible to obtain either with the 3-analogue-filter auditory prosthesis Triton 3004 from Siemens, or with our prototype as a function of the number of filters (3, 4 or 7) and their frequency band width programmability. We tested 21 patients suffering from middle or severe neurosensory hearing loss. This study allows to demonstrate that to overpass the present analogue T004 device a 7 programmable-width-filter strategy seems to be the most appropriate. Further studies benefiting with material improvement of our prototype and finer audiometric adjustment of filter strategies as well as long term clinical studies have to be carried out.

Auditory performances of a 3-4-7 programmable numeric filter hearing aid

We designed a non-portable prototype seven-filter digital auditory hearing aid. For each of the filters, frequency bandwidth, amplification and compression were programmable in order to adapt these parameters to the deaf patient's audiometric characteristics. We compared the hearing improvement it was possible to obtain either with the three-analogue filter auditory prosthesis Triton 3004 hearing aid from Siemens or with our prototype as a function of the number of filters (three, four or seven) and their frequency bandwidth programmability. We tested 21 patients suffering from moderate to severe sensorineural hearing loss. This study allowed us to demonstrate that a seven programmable-width filter strategy seems to be more effective than the present analogue T004 device. Further studies with improvement of our prototype and finer audiometric adjustment of filter strategies, together with long-term clinical studies, need to be carried out.

Asynchronous interleaved stimulation (AIS): a new speech coding strategy for cochlear implant

Cochlear implants use a fixed or F0 dependent stimulation rate, whatever the sound analysed (vowel or consonant) or procedure (fixed filters or FFT). We present a procedure which varies as a function of the input signal. This method is based on FFT analysis using a variable width analysis window. Short windows are applied on the transient part of the signal, providing poor frequency resolution but good time resolution; they permit us to accurately follow the speech signal in time during its fast temporal variations. Large windows are applied on the stationary parts of the signal, providing better frequency resolution but poor time resolution. Transient parts of the speech order this window switching; they are detected using the statistical properties of the FFT; moreover, the narrow windowing is coupled to an increase of the stimulation frequency. This strategy has been implemented using the DIGISONIC cochlear implant software, and clinically assessed in 6 regular cochlear implant users, on the basis of a consonant-vowel-consonant test. This Asynchronous-Interleaved-Stimulation (AIS) strategy provides the patients with better discrimination than fixed window FFT analysis. The coding protocol will be described and results presented.

Objective evaluation of vowel identification with the Digisonic cochlear implant

Perceptual assessment of speech recognition performance with cochlear implants is confronted with problems raised by patients' availability and specificity. These problems can be avoided by using objective evaluation methods based on automatic classification of speech sounds. In the present report, different tuning strategies for the Digisonic cochlear implant were evaluated by discriminant analysis. The results presented here concern the correct classification of vowels. A set of 600 vocalic items, corresponding to 10 vowel categories and produced by 20 French speakers, were processed by the Digisonic DX10 with two different strategies. In the linear strategy, center frequencies of the 15 Digisonic channels were almost equally spaced in acoustic units (Hz). In the mel strategy, center frequencies were almost equally spaced in psychoacoustic units. For each strategy, energy levels of the frequency channels were quantified for each vocalic item and processed by statistical discriminant analysis for classification into 10 vowel categories. Results mainly show that the percent correct classification (PCC) is larger for the mel strategy and that the improvement is due to a higher concentration of frequency channels below 4 kHz. Implications of these results for the tuning of cochlear implants are discussed.