EnSidNet: Enhanced Hybrid Siamese-Deep Network for grouping clinical trials into drug-development pathways
Lucia Pagani · 2021
Siamese Neural Networks have been widely used to perform similarity classification in multi-class settings.Their architecture can be used to group the clinical trials belonging to the same drugdevelopment pathway along the several clinical trial phases.Here we present an approach for the unmet need of drugdevelopment pathway reconstruction, based on an Enhanced hybrid Siamese-Deep Neural Network (EnSidNet).The proposed model demonstrates significant improvement above baselines in a 1-shot evaluation setting and in a classical similarity setting.EnSidNet can be an essential tool in a semi-supervised learning environment: by selecting clinical trials highly likely to belong to the same drug-development pathway it is possible to speed up the labelling process of human experts, allowing the check of a consistent volume of data, further used in the model's training dataset.38 of a class was seen by the algorithm only once 39 before making inference (Koch et al., 2015).40 Different architectures of SNN were developed in 41 time: Simo-Serra and colleagues developed a 3-42 inputs SNN (Simo-Serra et al., 2015), where the 43 neural network learned to rank the outputs and 44 identify whether the reference's hidden 45 representation is more similar to a positive or a 46 negative sample.47 Another example involves the insertion of an 48 intermediate stage between the similarity score 49 layer and the final prediction layer (Subramaniam, 50 Chatterjee, and Mittal, 2016), allowing to increase 51 performance in person re-identification task 52 despite partial occlusion and difference in point of 53 view or illumination.54 The first applications of SNN were based on 55 Convolutional Neural Networks (CNN) to obtain 56 similarity score on images (Simo-Serra et al., 57 2015), seeing SNN involved in different tasks 58 such as patch identification (Simo-Serra et al., 59 2015), person identification (Ahmed et al., 2015), 60 image matching from different angles (Vo and 61 Hays, 2016).SNN was also explored in Natural 62 Language Processing (NLP) contexts in tasks like 63 identifying sentence similarity (Mueller and 64 Thyagarajan, 2016) and support relation for 65 argumentation (Gema et al., 2017).These 66 applications highlight the flexibility of SNN to 67 identify similarities in different contexts.Here we 68 apply this architecture on an unmet healthcare 69 task: grouping clinical trials belonging to the same 70 drug-development pathway.71 Before being released on the market a new drug 72 needs to go through several expensive and time-73 consuming experiments, involving testing the 74 pharmacological characteristics of the drug in 75