應用深度神經網路進行不同兒童遺傳性聽損基因之預測

Using Deep Learning to Predict Audiological Outcomes for Different Genes of Hereditary Hearing Loss

曾苡珊(Yi-Shan Tseng)

指導教授 ： 周承復

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202202428

摘要

兒童聽損為一種遺傳性聽損，是臨床中常見的問題，也是在聽語中心相當重要的課題。本論文的目的為探索利用深度神經網路(Deep neural network)對不同的遺傳性聽損基因進行時間序列的分析，預測病患未來就診聽力值。提出的解決方案有：一種基於線性回歸的方式，另一種是基於深度學習的神經網路，如：多層感知機(Multilayer perceptron, MLP)、長短期記憶(Long Short-Term Memory, LSTM)與注意力機制(Attention)。另外，也提出一種內插法(Interpolation)來解決Irregularly sampled time series的問題，使得長短期記憶(Long Short-Term Memory, LSTM)與注意力機制(Attention)結合自定義的損失函示與正則項在預測結果中能有進步。而研究此議題之原因在於希望能幫助醫師能提前評估未來病人的就診聽力狀況，進而提早進行後續的治療，如：配戴助聽器、給予聽能復健訓練、植入人工耳蝸，甚至能提前開刀手術。如結果所示，運用神經網路在預測聽損的未來趨勢相當具有優勢，儘管醫院資料庫中病患就診資訊不一致或者不夠完整而具有挑戰性，但基於該結果我們可以得出一結論：利用深度神經網路，能準確且快速的預測未來就診聽力值與趨勢。

關鍵字

機器學習 ； 深度學習 ； 神經網路 ； 聽損 ； 聽損基因 ； 遺傳性聽損 ； GJB2 ； MYO15A

並列摘要

Hearing loss in children is a congenital or hereditary hearing loss, which is a common clinical problem and a very important topic in the listening and speaking center of hospitals in Taiwan. The purpose of this paper is to explore that using deep neural network for different genes of hearing loss to analyze the time series and predict the hearing values of patients in the future. The proposed solutions for this purpose: one based on linear regression, and the other based on deep neural networks, such as Fully connected networks, Long Short-Term Memory (LSTM), and Attention model. In addition, we proposed an interpolation method which is used to solve irregularly sampled time series, and also improved the prediction results of LSTM and Attention model. Similar to the interpolation method, it also called the data augmentation method in pre-processing. Finally, we use the customer loss function and regularization term to get better performance. The reason for studying this topic is to help doctors to assess the future hearing condition of patients in advance, and then to carry out follow-up treatments in advance, such as wearing hearing aids, giving hearing rehabilitation training, implanting cochlear implants, and even performing surgery in advance. Based on the results, using the neural networks is quite advantage in predicting the future trend of hearing loss. Although the patient information in the hospital database is inconsistent or incomplete are still challenging, we also concluded that our method is effective.

並列關鍵字

Machine learning ； Deep learning ； Neural network ； Hearing Loss ； Deafness gene ； GJB2 ； MYO15A

參考文獻

Chen PY, Lin YH, Liu TC, Lin YH, Tseng LH, Yang TH, Chen PL, Wu CC, Hsu CJ. Prediction Model for Audiological Outcomes in Patients With GJB2 Mutations. Ear Hear. 2020 Jan/Feb.

Chen PY, Tsai CY, Wu JL, Li YL, Wu CM, Chen KC, Hwang CF, Wu HP, Lin HC, Cheng YF, Lo MY, Liu TC, Yang TH, Chen PL, Hsu CJ, Wu CC. Hearing Features and Cochlear Implantation Outcomes in Patients With Pathogenic MYO15A Variants: a Multicenter Observational Study. Ear Hear. 2021 Dec 29.

Zhang, J., Guan, J., Wang, H. et al. Genotype-phenotype correlation analysis of MYO15A variants in autosomal recessive non-syndromic hearing loss. BMC Med Genet 20, 60 (2019).

Snoeckx RL, Huygen PL, Feldmann D, Marlin S, Denoyelle F, Waligora J, Mueller-Malesinska M, Pollak A, Ploski R, Murgia A, Orzan E, Castorina P, Ambrosetti U, Nowakowska-Szyrwinska E, Bal J, Wiszniewski W, Janecke AR, Nekahm-Heis D, Seeman P, Bendova O, Kenna MA, Frangulov A, Rehm HL, Tekin M, Incesulu A, Dahl HH, du Sart D, Jenkins L, Lucas D, Bitner-Glindzicz M, Avraham KB, Brownstein Z, del Castillo I, Moreno F, Blin N, Pfister M, Sziklai I, Toth T, Kelley PM, Cohn ES, Van Maldergem L, Hilbert P, Roux AF, Mondain M, Hoefsloot LH, Cremers CW, Löppönen T, Löppönen H, Parving A, Gronskov K, Schrijver I, Roberson J, Gualandi F, Martini A, Lina-Granade G, Pallares-Ruiz N, Correia C, Fialho G, Cryns K, Hilgert N, Van de Heyning P, Nishimura CJ, Smith RJ, Van Camp G. GJB2 mutations and degree of hearing loss: a multicenter study. Am J Hum Genet. 2005 Dec.

Zachary C Lipton, David Kale, and Randall Wetzel. Directly modeling missing data in sequences with rnns: Improved classification of clinical time series. In Machine Learning for Healthcare Conference, pp. 253–270, 2016.