A Multimodal Machine Learning Approach to Diagnosis, Prognosis, and Treatment Prediction for Neurodegenerative Diseases and Cancer

Despite early diagnosis and estimating the future course of neurodegenerative and cancerous diseases being integral for survival of patients, clinical and algorithmic methods fail to effectively utilize the multimodal data available, are time-inefficient, and expensive, making it difficult to access accurate screenings for these diseases. Therefore, a novel end-to-end quantum machine learning approach using multiple data modalities for the identification of diagnoses, prognoses, and effective treatments is proposed. In a procedural flow, data is sourced from one or more of the following: CT scan images, webcam, patient-physician audio, Whole Slide Images, and clinical data. For image data, a Convolutional Neural Network, is employed to detect high level features. With text-based clinical data (including audio-derived data), a Bidirectional Encoder Representation model is used to extract text embeddings. For video data, pupil progression and average fixation duration features are manually crafted. All feature vectors are concatenated, normalized, passed through a Deep Neural Network, and then mapped to one of 38 neurodegenerative and cancerous diseases. For prognosis, features are pooled, concatenated with the diagnosis feature vector, and passed through another neural network with an output of survival times. Treatment prediction involves an information-retrieval task matching feature vectors to treatments/drug lists from the FDA. The proposed approach was tested on 5,000 patient profiles sourced from the public TCGA and JPND databases, outperforming all other state-of-the-art approaches. The model predicted diagnoses with an accuracy of 98.53%, achieved a Concordance Index of 0.94 in predicting prognoses, and in treatment prediction achieved a 99.32% accuracy.