Identifying a cause of inaccuracy of a machine-learning algorithm
AssigneeKoninklijke Philips N.V.
InventorsDECLAN PATRICK KELLY, RIM HELAOUI
Life of this patent
Estimating the term…
Abstract
A mechanism for identifying contraindications for processing of subject data in a (data) record using a machine-learning method. Medical data in historic records are processed by the machine-learning method to identify predicted out-put data. The predicted output data is compared to corresponding data in the historic records to flag or otherwise identify historic records where there is a mismatch. The flagged historic records are then processed to identify any shared elements therebetween, which thereby represent contraindications for processing subject data using the machine-learning method.
Claims20
A computer-implemented method for identifying one or more potential causes of inaccuracy for a machine-learning method, the computer-implemented method comprising:
- obtaining a plurality of historic records, each associated with at least one of a different clinical subject or a subject at different points in time, each historic record containing subject data including:
- historic input medical data for input to the machine-learning method;
- and
- historic output medical data representing a ground-truth version of output medical data for the historic input medical data;
- for each historic record:
- processing a portion of the historic input medical data using the machine-learning method to produce predicted output medical data for the historic record;
- and
- flagging the historic record only when a difference between the predicted output medical data and the historic output medical data is determined to meet one or more predetermined criteria;
- and
- processing each flagged historic record to identify one or more elements of subject data common to more than a predetermined percentage or number of the subjects associated with the flagged historic records.
A non-transitory computer-readable storage medium having stored a computer program comprising instructions, which, when executed by a processor, cause the processor to:
- obtain a plurality of historic records, each associated with at least one of a different clinical subject or a subject at different points in time, each historic record containing subject data including:
- historic input medical data for input to the machine-learning method;
- and
- historic output medical data representing a ground-truth version of output medical data for the historic input medical data;
- for each historic record:
- process a portion of the historic input medical data using the machine-learning method to produce predicted output medical data for the historic record;
- and
- flag the historic record only when a difference between the predicted output medical data and the historic output medical data is determined to meet one or more predetermined criteria;
- and
- process each flagged historic record to identify one or more elements of subject data common to more than a predetermined percentage or number of subjects associated with the flagged historic records.
A system for identifying one or more potential causes of inaccuracy for a machine-learning method, the system comprising:
- a processor configured to:
- obtain a plurality of historic records, each associated with at least one of a different clinical subject or a subject at different points in time, each historic record containing subject data including:
- historic input medical data for input to the machine-learning method;
- and
- historic output medical data representing a ground-truth version of output medical data for the historic input medical data;
- for each historic record:
- process the historic input medical data using the machine-learning method to produce predicted output medical data for the historic record;
- and
- flag the historic record only when a difference between the predicted output medical data and the historic output medical data is determined to meet one or more predetermined criteria;
- and
- process each flagged historic record to identify one or more elements of subject data common to more than a predetermined percentage or number of subjects associated with the flagged historic records.
Dependencies are inferred from the claim text.