Comorbidity prediction from radiology images
Methods are provided to predict, based on chest radiographs, whether a patient suffers from any comorbidities in one or more high-level comorbidity classes, e.g., high-level hierarchical condition categories. The chest radiograph is a low-cost, minimally-invasive source of visual information that is able to accurately predict whether a subject suffers from diabetes with chronic complications, morbid obesity, congestive heart failure, specified heart arrhythmias, vascular disease, or chronic obstructive pulmonary disease, among other high-level comorbidity classes. The predictive methods provided herein are also able to accurately predict overall measures of health-related complications, including the risk adjustment factor. These predictive methods can be used to focus review of medical records, improving the detection of disease. Outputs generated by these methods can also be used to predict the severity of disease and/or the extent of care provided to subjects with COVID-19 or related diseases.
1. A computer-implemented method comprising:
receiving a target radiographic image of a chest of a subject, wherein the subject has a specified disease or syndrome, wherein an artificial neural network (ANN) has a set of inputs and a set of outputs, wherein the ANN has been trained by a training set including a plurality of radiographic images containing features that are mapped to the set of inputs and associated with labeled values mapped to the set of outputs, wherein a subset of the radiographic images contain instances of the features that are indicative of at least one comorbidity within a particular comorbidity class and are associated with instances of the labeled values that represent the particular comorbidity class, wherein the ANN a common layer and a plurality of parallel output heads, wherein the common layer provides outputs to the plurality of parallel output heads, and wherein each output head of the plurality of parallel output heads generates, as an output, a respective one of the set of outputs;
applying the ANN to the target radiographic image to generate a set of target outputs for the target radiographic image, wherein output values of the set of target outputs indicate that the subject has at least one comorbidity in the particular comorbidity class;
based on the set of target outputs for the target radiographic image, predicting at least one of (i) a degree of severity of the specified disease or syndrome or (ii) an amount of care that the subject will receive related to the specified disease or syndrome;
based on the set of target outputs for the target radiographic image, predicting a likelihood that the subject has a specified disease or syndrome;
determining that the likelihood that the subject has the specified disease or syndrome has an intermediate value; and
responsive to determining that the likelihood that the subject has the specified disease or syndrome has an intermediate value, performing an additional diagnostic test on the subject to determine whether or to what degree the subject has the specified disease or syndrome.
2. The method of claim 1 , wherein a further subset of the radiographic images contain no instances of the features that are indicative of any comorbidity within the particular comorbidity class and are not associated with any instances of the labeled values that represent the particular comorbidity class.
3. The method of claim 1 , wherein a subset of the radiographic images are associated with instances of the labeled values that represent one or more of (i) at least one comorbidity within a first comorbidity class that encompasses diabetes with chronic complications, (ii) at least one comorbidity within a second comorbidity class that encompasses morbid obesity, (iii) at least one comorbidity within a third comorbidity class that encompasses congestive heart failure, (iv) at least one comorbidity within a fourth comorbidity class that encompasses specified heart arrhythmias, (v) at least one comorbidity within a fifth comorbidity class that encompasses vascular disease, or (vi) at least one comorbidity within a sixth comorbidity class that encompasses chronic obstructive pulmonary disease.
4. The method of claim 1 , wherein a subset of the radiographic images are associated with instances of the labeled values that represent one or more of (i) at least one comorbidity within ICD10 hierarchical condition category HCC18, (ii) at least one comorbidity within ICD10 hierarchical condition category HCC22, (iii) at least one comorbidity within ICD10 hierarchical condition category HCC85, (iv) at least one comorbidity within ICD10 hierarchical condition category HCC96, (v) at least one comorbidity within ICD10 hierarchical condition category HCC108, or (vi) at least one comorbidity within ICD10 hierarchical condition category HCC111.
5. The method of claim 1 , wherein the labeled values mapped to the set of outputs include a numerical output that is indicative of an overall degree of medical risk.
6. The method of claim 5 , wherein the numerical output that is indicative of an overall degree of medical risk is a risk adjustment factor (RAF) score.
7. The method of claim 1 , wherein the labeled values mapped to the set of outputs represent, respectively, whether respective subjects represented in respective radiographic images of the plurality of radiographic images have: (i) at least one comorbidity within ICD10 hierarchical condition category HCC18, (ii) at least one comorbidity within ICD10 hierarchical condition category HCC22, (iii) at least one comorbidity within ICD10 hierarchical condition category HCC85, (iv) at least one comorbidity within ICD10 hierarchical condition category HCC96, (v) at least one comorbidity within ICD10 hierarchical condition category HCC108, or (vi) at least one comorbidity within ICD10 hierarchical condition category HCC111.
8. The method of claim 7 , wherein the labeled values mapped to the set of outputs additionally include a risk adjustment factor (RAF) score.
9. The method of claim 1 , further comprising:
applying a plurality of different modified versions of the target radiographic image to the ANN to generate a visual map that depicts a region within the target radiographic image that is indicative of the at least one comorbidity within the particular comorbidity class, wherein each one of the different modified versions of the target radiographic image has a respective different region of the target radiographic image blanked.
10. The method of claim 1 , further comprising:
applying the ANN to a particular radiographic image of the plurality of radiographic images of the training set to generate an updated set of outputs for particular radiographic image; and
updating the training set such that the particular radiographic image is associated with instances of the labeled values that reflect the updated set of outputs.
11. The method of claim 1 , further comprising:
receiving a pre-determined set of outputs for the target radiographic image, wherein output values of the pre-determined set of outputs for the target radiographic image indicate that the subject does not have at least one comorbidity within the particular comorbidity class;
determining that the pre-determined set of outputs for the target radiographic image differs from the set of target outputs generated by the ANN with respect to the subject having at least one comorbidity within the particular comorbidity class; and
responsively providing an indication via a user interface that the pre-determined set of outputs for the target radiographic image differs from the set of target outputs generated by the ANN.
12. The method of claim 1 , wherein the labeled values mapped to the set of outputs include a numerical output that is indicative of an overall degree of medical risk, and wherein the method additionally comprises:
receiving a pre-determined numerical indication of the overall degree of medical risk of the subject;
determining that the pre-determined numerical indication of the overall degree of medical risk of the subject differs by more than a specified threshold amount from the output, of the set of target outputs generated by the ANN, that is indicative of the overall degree of medical risk of the subject; and
responsively providing an indication via a user interface that the pre-determined numerical indication of the overall degree of medical risk of the subject differs from the output, of the set of target outputs generated by the ANN, that is indicative of the overall degree of medical risk of the subject.
13. The method of claim 1 , wherein predicting at least one of (i) a degree of severity of the specified disease or syndrome or (ii) an amount of care that the subject will receive related to the specified disease or syndrome comprises predicting at least one of: a likelihood of mortality or a likelihood of inpatient hospitalization longer than two days.
14. The method of claim 1 , wherein predicting at least one of (i) a degree of severity of the specified disease or syndrome or (ii) an amount of care that the subject will receive related to the specified disease or syndrome comprises applying a classifier to the set of target outputs for the target radiographic image.
15. The method of claim 14 , wherein the applying the classifier to the set of target outputs for the target radiographic image comprises applying a generalized linear model to the set of target outputs for the target radiographic image.
16. The method of claim 14 , wherein applying the classifier to the set of target outputs for the target radiographic image comprises determining a weighted sum of the set of target outputs for the target radiographic image and comparing the weighted sum to a pre-specified threshold value.
17. The method of claim 1 , further comprising:
applying an additional ANN to the target radiographic image to generate one or more additional outputs, wherein the additional ANN has been trained based on an additional training set that includes a plurality of training examples from people determined to have the specified disease or syndrome, and wherein predicting at least one of (i) a degree of severity of the specified disease or syndrome or (ii) an amount of care that the subject will receive related to the specified disease or syndrome based on the set of target outputs for the target radiographic image comprises predicting at least one of (i) a degree of severity of the specified disease or syndrome or (ii) an amount of care that the subject will receive related to the specified disease or syndrome based on the set of target outputs for the target radiographic image and the one or more additional outputs.
18. The method of claim 17 , wherein the one or more additional outputs of the additional ANN are indicative of at least one of a severity of lung opacity, a degree of lung opacity, or an extent of lung opacity depicted in the target radiographic image.
19. The method of claim 18 , wherein the specified disease or syndrome is SARS-CoV-2 or a variant thereof.
20. A non-transitory computer-readable medium, having stored thereon program instructions that, upon execution by a computing device, cause the computing device to perform a method comprising:
receiving a target radiographic image of a chest of a subject, wherein the subject has a specified disease or syndrome, wherein an artificial neural network (ANN) has a set of inputs and a set of outputs, wherein the ANN has been trained by a training set including a plurality of radiographic images containing features that are mapped to the set of inputs and associated with labeled values mapped to the set of outputs, wherein a subset of the radiographic images contain instances of the features that are indicative of at least one comorbidity within a particular comorbidity class and are associated with instances of the labeled values that represent the particular comorbidity class, wherein the ANN a common layer and a plurality of parallel output heads, wherein the common layer provides outputs to the plurality of parallel output heads, and wherein each output head of the plurality of parallel output heads generates, as an output, a respective one of the set of outputs;
applying the ANN to the target radiographic image to generate a set of target outputs for the target radiographic image, wherein output values of the set of target outputs indicate that the subject has at least one comorbidity in the particular comorbidity class;
based on the set of target outputs for the target radiographic image, predicting at least one of (i) a degree of severity of the specified disease or syndrome or (ii) an amount of care that the subject will receive related to the specified disease or syndrome;
based on the set of target outputs for the target radiographic image, predicting a likelihood that the subject has a specified disease or syndrome;
determining that the likelihood that the subject has the specified disease or syndrome has an intermediate value; and
responsive to determining that the likelihood that the subject has the specified disease or syndrome has an intermediate value, performing an additional diagnostic test on the subject to determine whether or to what degree the subject has the specified disease or syndrome.