	US 12,431,247 B2
	Systems and methods for prognosis prediction of acute myeloid leukemia patients
Bor-Sheng Ko, Taipei (TW); Yu-Fen Wang, Taipei (TW); Chi-Chun Lee, Hsinchu (TW); Jeng-Lin Li, Hsinchu (TW); and Jih-Luh Tang, Taipei (TW)
Assigned to National Taiwan University, Taipei (TW)
Filed by National Taiwan University, Taipei (TW)
Filed on Mar. 11, 2020, as Appl. No. 16/815,017.
Prior Publication US 2021/0287805 A1, Sep. 16, 2021
Int. Cl. G16H 50/30 (2018.01); G06N 3/08 (2023.01); G06N 20/10 (2019.01); G16H 10/40 (2018.01); G16H 10/60 (2018.01)

CPC G16H 50/30 (2018.01) [G06N 3/08 (2013.01); G06N 20/10 (2019.01); G16H 10/40 (2018.01); G16H 10/60 (2018.01)]

19 Claims

1. A computer-implemented method for predicting prognosis of a patient that has been diagnosed with acute myeloid leukemia, the method comprising:

receiving data that is associated with the patient and that comprises

(i) a first static attribute that is representative of an age or a gender of the patient at diagnosis,

(ii) a second static attribute that is representative of a categorization for acute myeloid leukemia determined as a result of cytogenetic testing performed at, or subsequent to, diagnosis, and

(iii) a time-dependent attribute represented as a multi-dimensional vector that indicates, for each of a plurality of genes, whether that gene has mutated or not at each of a plurality of time windows that collectively define an interval of time;

processing the time-dependent attribute using a trained neural network such that for each of the plurality of time windows, the time-dependent attribute is encoded into a different one of a first plurality of vectors, each of which serves as a time-series representation of a corresponding one of the plurality of time windows;

converting the first static attribute into a first static variable and the second static attribute into a second static variable;

concatenating the first plurality of vectors and the first and second static variables to produce a second plurality of vectors, each of which includes (i) a different one of the first plurality of vectors, (ii) the first static variable, and (iii) the second static variable; and

producing a prognosis outcome for the patient by providing the second plurality of vectors to a trained classifier as input, such that the trained classifier considers the first and second static attributes in combination with each one of the first plurality of vectors in producing the prognosis outcome;

wherein the trained classifier is trained by steps of

(a) assembling a training data set comprising a retrospective collection of data of multiple patients that are known to have acute myeloid leukemia, wherein the retrospective collection of data comprises a collected number of static attributes,

time-dependent attributes and mortality and relapse outcomes of the multiple patients;

(b) processing the time-dependent attributes of the training data set using the trained neural network to encode the time-dependent attributes into vectors that serve as time-series representations;

(d) combining the time-series representations and the static variables to train a classifier based on the combined time-series representations and the static variables.

15. A system for predicting prognosis of a patient that has been diagnosed with acute myeloid leukemia, the system comprising:

at least one processor operatively coupled with a datastore, the at least one processor configured to:

receive, from a file storage means, data that is associated with the patient and that comprises

(i) a first static attribute that is representative of an age or a gender of the patient at diagnosis,

(ii) a second static attribute that is representative of a categorization for acute myeloid leukemia determined as a result of cytogenetic testing performed at, or subsequent to, diagnosis, and

(iii) a time-dependent attribute that is represented as a multi-dimensional vector that indicates, for each of a plurality of genes, whether that gene has mutated or not at each of a plurality of time windows that collectively define an interval of time;

process the time-dependent attribute using a trained neural network such that for each of the plurality of time windows, the time-dependent attribute is encoded into a different one of a first plurality of vectors, each of which serves as a time-series representation of a corresponding one of the plurality of time windows;

convert the first static attribute into a first static variable and the second static attribute into a second static variable;

combine the first plurality of vectors and the first and second static variables to produce a second plurality of vectors, each of which includes (i) a different one of the first plurality of vectors, (ii) the first static variable, and (iii) the second static variable; and

provide a prognosis outcome for the patient on a display means by providing the second plurality of vectors to a trained classifier as input;

wherein the trained classifier is trained by steps of

(a) assembling a training data set comprising a retrospective collection of data of multiple patients that are known to have acute myeloid leukemia from a database, wherein the retrospective collection of data comprises a collected number of static attributes, time-dependent attributes and mortality and relapse outcomes of the multiple patients;

(b) processing the time-dependent attributes of the training data set using the trained neural network to encode the time-dependent attributes into vectors that serve as time-series representations;

(d) combining the time-series representations and the static variables to train a classifier based on the combined time-series representations and the static variables.