Apparatus and method for training a machine learning model
Inventors
Aravamudan, Murali • Rajasekharan, Ajit
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12259864-B1
Publication Date
2025-03-25
Expiration Date
Abstract
Described herein is an apparatus and method for training a machine learning model. An apparatus may include a computing device configured to receive a corpus of data containing entries corresponding to a plurality of subjects; identify a plurality of entries within the corpus, corresponding to a first subject of the plurality of subjects and representing medical history of the first subject; determine a plurality of temporal attributes of the plurality of entries; generate a plurality of tokens as a function of the plurality of entries; generate a chronological data structure segment ordering one or more of the plurality of entries and the plurality of tokens, as a function of the plurality of temporal attributes; and train a multimodal machine learning model on a training dataset including the chronological data structure segment.
Core Innovation
The invention provides an apparatus for training a machine learning model using medical-history data represented by entries corresponding to a plurality of subjects. The apparatus receives, from databases via a network, a corpus of data containing the entries, extracts a plurality of entries with an identifier corresponding to a first subject, and represents medical history of the first subject. It then executes a temporal attribute machine learning model to determine temporal attributes of the entries, where at least a temporal attribute represents time within the medical history.
Determining the plurality of temporal attributes comprises training the temporal attribute machine learning model on a training dataset correlated to the temporal attributes, and generating a temporal attribute as a function of the entry using the trained temporal attribute machine learning model. The invention further alters the temporal attribute while preserving the temporal attribute's chronological order relative to other temporal attributes of entries having a same subject. Based on the determined temporal attributes, the apparatus generates automatically a plurality of tokens as a function of the entries, wherein the plurality of tokens comprises tokens of different modalities.
The apparatus generates a chronological data structure segment comprising the entries and the plurality of tokens as a function of the temporal attributes. It trains a multimodal machine learning model on a training dataset including the chronological data structure segment to produce model output including a medical prediction as a medical state of the first subject. The multimodal training comprises computing co-occurrences of the tokens of different modalities to compute degree similarity for the tokens.
Claims Coverage
The independent claims are clm-00001 and clm-00010. Across these, the main inventive features include temporal-attribute determination with chronological-order-preserving alteration, automatic generation of different-modality tokens, construction of a chronological data structure segment, and multimodal training producing a medical prediction using token co-occurrences and degree similarity.
Receiving a corpus of medical-history entries from databases via a network
Receive, from databases via a network, a corpus of data containing entries corresponding to a plurality of subjects.
Extracting entries for a first subject representing medical history
Extract a plurality of entries with an identifier corresponding to a first subject and represents medical history of the first subject.
Temporal attribute machine learning model to determine temporal attributes representing time
Execute a temporal attribute machine learning model to determine a plurality of temporal attributes of the plurality of entries, wherein at least a temporal attribute represents time within the medical history.
Temporal attribute model training and generation from entries
Train the temporal attribute machine learning model on a training dataset including the plurality of entries correlated to the plurality of temporal attributes, and generate a temporal attribute as a function of the entry using the trained temporal attribute machine learning model.
Altering temporal attributes while preserving chronological order within a subject
Alter the temporal attribute while preserving the temporal attribute's chronological order relative to other temporal attributes of entries having a same subject.
Automatic generation of multi-modality tokens from determined temporal attributes
Generate automatically, based on the determined plurality of temporal attributes, a plurality of tokens as a function of the plurality of entries, wherein the plurality of tokens comprises tokens of different modalities.
Chronological data structure segment from entries and tokens as a function of temporal attributes
Generate a chronological data structure segment comprising the plurality of entries and the plurality of tokens, as a function of the plurality of temporal attributes.
Multimodal machine learning model training to produce medical prediction
Train a multimodal machine learning model on a training dataset including the chronological data structure segment to produce model output including a medical prediction as a medical state of the first subject.
Computing co-occurrences of tokens to compute degree similarity for different modalities
During the training, compute co-occurrences of the tokens of different modalities to compute degree similarity for the tokens of different modalities.
Receiving model output from the trained multimodal model
Receive, from the trained multimodal machine learning model run on a computing device, the model output.
Both independent claims define a pipeline that receives a corpus of subject medical-history entries, extracts entries for a first subject, determines temporal attributes using a trained temporal attribute machine learning model with chronological-order-preserving alteration, generates automatically multi-modality tokens, builds a chronological data structure segment, and trains a multimodal machine learning model using token co-occurrences to compute degree similarity, producing model output including a medical prediction.
Stated Advantages
Documented Applications
No documented applications found
Interested in licensing this patent?