Privacy preserving artificial intelligence based clinical decision support

1. A method for performing machine-learning tasks on medical data, the method comprising:

receiving, by a processor, medical data;

encoding, by the processor, the medical data into encoded medical data according to an encoding scheme, wherein the encoding scheme represents a number in the medical data as a series of polynomial terms, wherein each polynomial term comprises a sub-unity number and a negative power of ten;

encrypting, by the processor, the encoded medical data into encrypted medical data according to a homomorphic encryption scheme;

applying, by the processor, the encrypted medical data to a machine-learned model, the machine learned model trained on second medical data;

receiving, by the processor, an encrypted result generated by the machine-learned model based on applying the encrypted medical data to the machine-learned model;

decrypting, by the processor, the encrypted result to an encoded result according to the homomorphic encryption scheme;

decoding, by the processor, the encoded result to a medical data result according to the encoding scheme; and

outputting, by the processor, the medical data result.

2. The method of claim 1 , wherein the machine-learned model is cloud-based.

3. The method of claim 1 , wherein a rational number is represented as the series of polynomial terms by the encoding scheme.

4. The method of claim 1 , wherein negative numbers in the medical data are represented by a difference between two polynomial terms.

5. The method of claim 1 , wherein null digits of the number are unencoded by the encoding scheme.

6. The method of claim 1 , further comprising:

receiving, by the processor, a precision parameter,

wherein a number of the series of polynomial terms is based on the precision parameter.

7. The method of claim 1 , wherein the medical data is encrypted with a cryptographic key used to encrypt the second medical data.

8. A method of training a machine-learning model, the method comprising:

receiving, by a processor, training data comprising medical data and result data, wherein the medical data is annotated with corresponding result data;

encoding, by the processor, the training into encoded training data according to an encoding scheme, wherein the encoding scheme represents a rational number as a series of polynomial terms, wherein each polynomial term comprises a sub-unity number and a negative power of ten;

encrypting, by the processor, the encoded training data into encrypted training data according to a homomorphic encryption scheme;

training with machine learning, by the processor, the machine-learning model based on the encrypted training data, wherein a result of the training is a machine-learned model configured to accept as input medical data and output result data, and wherein training further comprises:

generating, by the processor, two matrices based on the encrypted training data;

generating, by the processor, parameters of the machine-learning model based on the two matrices and a cost function comprising a linear system; and

storing, by the processor, the machine-learned model.

9. The method of claim 8 , further comprising:

encoding, by the processor, the parameters of the machine-learning model into encoded parameters,

wherein the machine-learned model is stored with the encoded parameters.

10. The method of claim 9 , wherein the homomorphic encryption scheme includes processing scalar operations.

11. The method of claim 9 , wherein a result of the training is a machine-learned model configured to accept as input encrypted medical data and output encrypted result data.

12. The method of claim 9 , wherein the training data is aggregated into a matrix form.

13. The method of claim 8 , further comprising:

receiving, by a processor, the medical data;

encoding, by the processor, the medical data into encoded medical data according to the encoding scheme;

encrypting, by the processor, the encoded medical data into encrypted medical data according to a homomorphic encryption scheme;

applying, by the processor, the encrypted medical data to the machine-learned model;

receiving, by the processor, encrypted result generated by the machine-learned model based on the encrypted medical data sent to the machine-learned model;

decrypting, by the processor, the encrypted result to an encoded result according to the homomorphic encryption scheme;

decoding, by the processor, the encoded result to a medical data result according to the encoding scheme; and

outputting, by the processor, the medical data result.

14. The method of claim 13 , wherein parameters of the machine-learned model, the encrypted medical data, and the encrypted training data are encrypted with the same encryption key.

15. The method of claim 13 , wherein parameters of the machine-learned model are encoded according to the encoding scheme,

wherein the encrypted medical data is encrypted with a first key, and

wherein the encrypted training data is encrypted with a second key different from the first key.

16. A clinical processing system, the system comprising: a processor, coupled with a memory containing instructions that, when executed, cause the processor to: send encoded-encrypted numerical data to a machine-learned model, wherein the encoded-encrypted numerical data is encoded according to an encoding scheme, wherein the encoding scheme represents a rational number as a series of polynomial terms, wherein each polynomial term comprises a sub-unity number and a negative power of ten; receive an encoded-encrypted result generated by the machine-learned model based on the encoded-encrypted numerical data sent to the machine-learned model; decrypt the encoded-encrypted result to an encoded result according to a homomorphic encryption scheme; decode the encoded result to a numerical data result according to the encoding scheme; and output the numerical data result to the memory.