Can we ask computers to recognize what we see from brain signals alone? Our paper seeks to utilize the knowledge learnt in the visual domain by popular pre-trained vision models and use it to teach a recurrent model being trained on brain signals to learn a discriminative manifold of the human brain's cognition of different visual object categories in response to perceived visual cues. For this we make use of brain EEG signals triggered from visual stimuli like images and leverage the natural synchronization between images and their corresponding brain signals to learn a novel representation of the cognitive feature space. The concept of knowledge distillation has been used here for training the deep cognition model, CogniNet 1 , by employing a student-teacher learning technique in order to bridge the process of inter-modal knowledge transfer. The proposed novel architecture obtains stateof-the-art results, significantly surpassing other existing models. The experiments performed by us also suggest that if visual stimuli information like brain EEG signals can be gathered on a large scale, then that would help to obtain a better understanding of the largely unexplored domain of human brain cognition.

No abstract

e14056 Background: Determination of the metastatic status of a patient is important for outcomes research and candidacy for clinical trials. Structured data in EMR may not always capture the metastatic status, and it is useful to extract it automatically from physician notes. Contextual understanding of the notes is important to resolve issues such as a) local vs distal metastasis b) statements involving family history of metastasis or physician instructing the patient to look for certain signs of metastasis c) text indicating suspicion of metastasis or absence of metastasis d) indirect utterances, e.g. cancer has spread to the bone. e) corrections to previous findings. Methods: We used a set of 20138 breast cancer patients from Concerto HealthAI real world oncology dataset that includes data from CancerLinQ Discovery to build & validate the set of NLP algorithms. 5300 sentences from 1500 patients were annotated & algorithms manually validated by data abstractors for 500 patients. The algorithms developed were the following: 1) Classification of a sentence into 3 classes: Distal/Local metastasis, Suspicious & Other 2) Classification of a sentence into 2 classes: Distal or Local 3) Classification of a patient into 2 classes: Distal metastasis or not distal metastasis 4) Multi label classification for detecting sites of metastasis. Sentence level algorithms were built using Deep Learning and patient level aggregation of sentence level prediction was done using ML approaches including temporal features. Pretrained ULMFiT model was fine-tuned with Concerto HealthAI’s corpus for sentence classification tasks. Results: At a sentence level, we obtained an accuracy of 0.85 for the distal/local vs suspicious vs irrelevant model and 0.97 for the distal vs not distal metastasis model. Our patient level metrics are shown in the table. The classes used for sites of metastasis are Brain, Bone, Lung, Liver, Distant Lymph nodes & Unknown sites. Subset accuracy (mean fraction of labels which match ) of 0.93 was obtained on the hold out test set at patient level. Conclusions: Metastatic status & site of metastasis can be reliably extracted automatically from clinical notes using deep learning techniques. This information will be valuable for clinical trial matching, outcomes research and other applications. [Table: see text]

No abstract