Machine learning techniques using cross-model fingerprints for novel predictive tasks

A system comprising:

• one or more processors;
• and
• at least one memory storing processor-executable instructions that, when executed by the one or more processors, cause the one or more processors to:
• input a predictive input to at least one first task prediction machine learning model trained to perform at least one first predictive task to receive an affirmative fingerprint distance measure, wherein:
• (i) the affirmative fingerprint distance measure is associated with a second predictive task for which the at least one first task prediction machine learning model is not trained,
• (ii) the second predictive task corresponds to a limited training dataset with less training data relative to one or more training datasets corresponding to the at least one first predictive task, and
• (iii) determining the affirmative fingerprint distance measure comprises:
• generating, based at least in part on the predictive input and using the at least one first task prediction machine learning model, a plurality of per-model prediction scores for the predictive input,
• determining a first cross-model prediction score for the predictive input based at least in part on the plurality of per-model prediction scores,
• generating a plurality of affirmative-labeled per-model prediction scores corresponding to a set of affirmative-labeled predictive inputs by providing the set of affirmative-labeled predictive inputs as input to the at least one first task prediction machine learning model, wherein an affirmative-labeled predictive input of the set of affirmative-labeled predictive inputs comprises a predictive input associated with an affirmative ground-truth label with respect to the second predictive task,
• determining an affirmative cross-model fingerprint for the second predictive task based on combining the plurality of affirmative-labeled per-model prediction scores, and
• determining, based at least in part on a vector representation of the first cross-model prediction score and a vector representation of the affirmative cross-model fingerprint, an affirmative fingerprint distance measure for the predictive input;
• and
• input the predictive input and the affirmative fingerprint distance measure to a second task prediction machine learning model configured to generate a predictive output indicating a likelihood of a rare disease.

One or more non-transitory computer-readable storage media including instructions that, when executed by one or more processors, cause the one or more processors to:

• input a predictive input to at least one first task prediction machine learning model, trained to perform at least one first predictive task, to receive an affirmative fingerprint distance measure, wherein:
• (i) the affirmative fingerprint distance measure is associated with a second predictive task for which the at least one first task prediction machine learning model is not trained,
• (ii) the second predictive task corresponds to a limited training dataset with less training data relative to one or more training datasets corresponding to the at least one first predictive task, and
• (iii) determining the affirmative fingerprint distance measure comprises:
• generating, based at least in part on the predictive input and using the at least one first task prediction machine learning model, a plurality of per-model prediction scores for the predictive input,
• determining a first cross-model prediction score for the predictive input based at least in part on combining the plurality of per-model prediction scores,
• generating a plurality of affirmative-labeled per-model prediction scores corresponding to a set of affirmative-labeled predictive inputs by providing the set of affirmative-labeled predictive inputs as input to the at least one first task prediction machine learning model, wherein an affirmative-labeled predictive input of the set of affirmative-labeled predictive inputs comprises a predictive input associated with an affirmative ground-truth label with respect to the second predictive task,
• determining an affirmative cross-model fingerprint for the second predictive task based on the plurality of affirmative-labeled per-model prediction scores, and
• determining, based at least in part on a vector representation of the first cross-model prediction score and a vector representation of the affirmative cross-model fingerprint, an affirmative fingerprint distance measure for the predictive input;
• and
• input the predictive input and the affirmative fingerprint distance measure to a second task prediction machine learning model configured to generate a predictive output indicating a likelihood of a rare disease.