Personalized health risk assessment for critical care

A computer-implemented method of assessing whether a patient is at risk of an adverse outcome relating to one or more clinical conditions, the method comprising:

• receiving training data and target data, wherein
• receiving training data includes receiving data x1 through xn, each xi of x1 through xn having one dimension for each of a plurality of patient characteristics and indicating values of the plurality of patient characteristics for an ith patient of n patients, and
• receiving target data includes receiving data xtp having one dimension for each of the plurality of patient characteristics and indicating values of the plurality of patient characteristics for a target patient;
• determining via the computer whether, or an extent to which, xtp is anomalous, wherein determining whether, or the extent to which, xtp is anomalous includes
• (a) for each xi of x1 through xn, calculating a distance between xi and k nearest neighbors of xi ,
• (b) calculating a distance between xtp and k nearest neighbors of xtp,
• (c) determining a threshold based at least on the distances calculated in (a), and
• (d) determining whether, or the extent to which, xtp is anomalous at least in part by comparing the distance calculated in (b) to the threshold,
• wherein calculating the distances in (a) and (b) includes, for at least some of the distances calculated in (a) and (b), finding the respective k nearest neighbors by implementing a plurality of locality sensitive hashing (LSH) searches with a progressively increasing radius, the radius being increased at least until the respective k nearest neighbors are found;
• and
• indicating whether, or an extent to which, the target patient is at risk of the adverse outcome, wherein the indication is based at least in part on the determination of whether, or the extent to which, xtp is anomalous.

A computer-implemented method of assessing whether a patient is at risk of developing a clinical condition, the method comprising:

• receiving training data and target data, wherein
• receiving training data includes receiving data sets (x1, y1) through (xn, yn), each xiof x1 through xn having one dimension for each of a plurality of patient characteristics and indicating values of the plurality of patient characteristics for an ith patient of n patients, and each yi of y1 through yn indicating whether the ith patient of the n patients is positively associated with the clinical condition, and
• receiving target data includes receiving data xtp having one dimension for each of the plurality of patient characteristics and indicating values of the plurality of patient characteristics for a target patient;
• generating model data via a computer using the training data, wherein generating the model data includes
• minimizing an objective function by selecting a multidimensional vector w and variables ψi, ξi, b and p subject to constraints that include
• wTΦ(xi)−p ≧(−ψi) for all i from 1 to n,
• yi(wTΦ(xi)−b)≧1−ξi for all i from 1 to n,
• ψi≧0, and
• ξi≧0,
• wherein the objective function is
• 1
• 2
• ⁢
• 
• w
• 
• 2
• -
• ρ
• +
• C
• 1
• ⁢
• ∑
• i
• =
• 1
• n
• ⁢
• ⁢
• ψ
• i
• +
• C
• 2
• ⁢
• ∑
• i
• =
• 1
• n
• ⁢
• ⁢
• ξ
• i
• ,
• and
• wherein C1 and C2 are costs assigned to penalties for a 1-class support vector machine (SVM) problem and a 2-class SVM problem, respectively;
• and
• at least one of
• (a) determining via the computer a classification for the target data, or
• (b) determining via the computer a risk score for the target data,
• wherein determining the classification for the target data or determining the risk score for the target data includes performing one or more calculations involving the target data and the model data, and
• wherein performing the one or more calculations includes comparing xtp to a decision boundary that is a function of the multidimensional vector w;
• and
• at least one of
• (a) indicating whether the target patient is at risk of developing the clinical condition, wherein the indication is based at least in part on the classification, or
• (b) indicating an extent to which the target patient is at risk of developing the clinical condition, wherein the indication is based at least in part on the risk score.