Patient diabetes monitoring system with clustering of unsupervised daily CGM profiles (or insulin profiles) and method thereof

• US 10,575,790 B2
• Filed: 03/02/2016
• Issued: 03/03/2020
• Est. Priority Date: 03/02/2016
• Status: Active Grant

• Current Assignee: Roche Diabetes Care Inc. (Roche Holding AG)
• Original Assignee: Roche Diabetes Care Inc. (Roche Holding AG)
• Inventors: Duke, David L., Steiger, Bernd, Manohar, Chinmay Uday
• Primary Examiner(s): Weare, Meredith
• Application Number:
  US15/058,271
• Publication Number:
  US 20170251980A1
• Time in Patent Office: 1,462 Days
• CPC Class Codes: A61B 5/14503 invasive, e.g. introduced into the body by a catheter or needle or using implanted sensors A61B5/1459, A61B5/1464, A61B5/1473, A61B5/1482, A61B5/14865 take precedence
• CPC Class Codes: A61B 5/14532 for measuring glucose, e.g. by tissue impedance measurement
• CPC Class Codes: A61B 5/4842 Monitoring progression or stage of a disease
• CPC Class Codes: A61B 5/7246 using correlation, e.g. template matching or determination of similarity
• CPC Class Codes: A61B 5/7271 Specific aspects of physiological measurement analysis A61B5/316, A61B5/349 and A61B5/374 take precedence
• CPC Class Codes: A61B 5/742 using visual displays A61B5/339 takes precedence
• CPC Class Codes: G06F 17/16 Matrix or vector computation , e.g. matrix-matrix or matrix-vector multiplication, matrix factorization matrix transposition G06F7/78
• CPC Class Codes: G06F 18/23 Clustering techniques
• CPC Class Codes: G06F 2218/00 Aspects of pattern recognition specially adapted for signal processing
• CPC Class Codes: G16H 20/17 delivered via infusion or injection

Abstract

A patient diabetes monitoring system with an efficient unsupervised daily monitoring profile clustering algorithm, a method, and a computer product thereof are disclosed. The system may include a physiological data input device or sensor which receives a plurality of physiological measurements to generate a dataset, a memory which stores a clustering algorithm, and a processor. The clustering algorithm when executed by the processor, causes the processor to automatically pre-process the dataset to control an amount of bias/aggressiveness from the collected unsupervised daily monitoring profiles, thereby generating a pre-processed dataset, build a similarity matrix from the pre-processed dataset, and output an optimum number of similarity clusters found by the processor from the similarity matrix.

Claims

1. A patient diabetes monitoring system for a patient comprising:

a physiological data input device which acquires a plurality of blood glucose concentration measurements of the patient within a time window to generate at least one time window dataset of collected unsupervised daily monitoring profiles;

a memory storing an unsupervised daily monitoring profile clustering algorithm; and

a processor in communication with said input device to receive said generated at least one time window dataset, and in communication with said memory in order to execute said unsupervised daily monitoring profile clustering algorithm,

wherein said unsupervised daily monitoring profile clustering algorithm when executed by said processor causes said processor automatically to;

pre-process the dataset to generate a pre-processed dataset via a data transformation of the dataset that makes the pre-processed dataset symmetric for retrospective analysis,

build a similarity matrix from the pre-processed dataset, and

output an optimum number of similarity clusters found by the processor from the similarity matrix,

wherein the data transformation for retrospective analysis results from processing the dataset with a hazard function defined by;

G_t=α

*ln(G−

β

)−

α

*ln(α

), where parameter α

=T_c−

β

, and parameter β

=D_r−

1, in which T_c is a center of a transformed space, D_r is a minimum defined glucose level, G_t is the transformed data of the blood glucose concentration measurements provided in the dataset, and G is original glucose level values of the blood glucose concentration measurements provided in the dataset and measured in millimoles per liter.

2. The system of claim 1, wherein after the pre-processing of the dataset, the pre-processed dataset is then processed to build the similarity matrix to account for time-series dynamics in the pre-processed dataset.

3. The system of claim 2, wherein the time-series dynamics in the pre-processed dataset is accounted for by a distance matrix that accounts for glucose value levels in an actual space or transformed space as well as via a rate of change of the glucose value levels to compute a distance between each pair of similar time series of data presented in the pre-processed dataset.

4. The system of claim 3, wherein the distance matrix is defined by:

d(X_i,Y_i)=k*|X_i−

Y_i|+(1−

k)*|(m_x−

m_y)*(X_i−

Y_i)|,

where, X_i is a glucose level value in a first time series X at time i, Y_i is a glucose value in a second time series Y at time i, k is a weighing factor, m_x is the slope at time i for the first time series X_i, and m_y is the slope at time i for time series X_i.

5. The system of claim 4, wherein a sum of distances between the first and second time series X and Y is used in an elastic alignment procedure to account for varying temporal responses/shifts in the pre-processed dataset.

6. The system of claim 5, wherein the elastic alignment procedure is a dynamic time warping process which allows for elastic matching of the first and second time series X and Y by local compression or elongation along a time axis.

7. The system of claim 6, wherein the dynamic time warping process results in any penalty being added to the sum of the distances between the first and second time series X and Y.

8. The system of claim 7, wherein the first and second time series are continuous glucose monitor (CGM) curves.

9. The system of claim 7, wherein the first and second time series X and Y of the pre-processed dataset are processed by the processor with the penalty as follows:

(a) Start at origin, distance between curves of the first time series X and the second time series Y is;

X(1,1)=Y(1,1);

(b) Keep first row a constant distance by;

X(i,1)=X(i−

1,1)+Y(i,1);

(c) Keep first column constant by;

X(1,j)=X(1,j−

1)+Y(1,j); and

(d) Carry on for next row and next column to end of search space of the pre-processed dataset as defined by;

X(i, j)=min(X(i, j−

1), X(i−

1, j−

1), X(i−

1, j))+Y(i, j).

10. The system of claim 1, wherein output of the build a similarity matrix process is checked against one or more conditions to evaluate if a determined alignment path is a valid path, the one or more conditions being:

monotonicity, continuity, boundary conditions, search window, and slope.

11. The system of claim 10, wherein output of the similarity matrix process is then used in an agglomerative clustering process to output similarity clusters, the agglomerative clustering process having the following pseudo code:

(a) Compute a distance matrix between data points of the output;

(b) Let each of the data points be a cluster;

(c) Repeat following;

i. Merge two closest clusters, and

ii. Update the distance matrix; and

(d) Do Repeat until only a single cluster remains.

12. The system of claim 11, wherein an inflection point in the distance matrix is calculated by the processor to find the optimal minimum number of clusters.

13. The system of claim 12, wherein if d(l) is a distance curve in the distance matrix, d′

(l) is a first derivative of the distance curve, and d″

(l) is a second derivative of the distance curve, and if d′

(l) exists, then the optimal minimum number of clusters along the curve d(l) is calculated by the processor to be a point l where d″

(l)=0.

14. The system of claim 12, wherein the processor calculates the inflection point as follows:

(a) Let first k points on the distance curve d(l) with p points be 1, 2, . . . k, and find slopes;

m₁=d(2)−

d(1)/(2−

1),m₂=d(3)−

d(1)/(3−

1), . . . ,m_k=d(n)−

d(1)/(n−

1);

(b) Calculate median of slopes from step (α

);

m_a=median(m₁, m₂ . . . m_k);

(c) Let last n points on the distance curve d(l) with p points be p−

n, . . . , p−

1, p, and find slopes;

m_p=d(p)−

d(p−

1)/(p−

(p−

1)), m₂=d(p)−

d(p−

2)/(p−

(p−

2)), . . . , m_n=d(p)−

d(p−

n)/(p−

(p−

n)); and

(d) Calculate median of slopes from step (c);

m_b=median(m₁, m₂ . . . m_n),

where a first line defined by the median slope m_a with a starting point as the first point along the distance curve d(l), and second line being defined by the median slope m_b with a starting point as the end point along the distance curve d(l), the inflection point being a projection of an intersection point between the first and second lines on the distance curve d(l) denoted by l_p, and if inflection point l_p is not an integer, then the optimal minimum number of clusters L_min is found by;

15. The system of claim 1, wherein the physiological data input device is a continuous glucose monitor (CGM).

16. A non-transitory computer-readable medium that stores a program that, when executed by a processor, causes the processor to execute, via a patient diabetes monitoring system having a physiological data input device which acquires a plurality of blood glucose concentration measurements of the patient within a time window to generate at least one time window dataset of collected unsupervised daily monitoring profiles and which is in communication with said processor, such that said processor receives said generated at least one time window dataset, and in communication with said memory, an unsupervised daily monitoring profile clustering algorithm that causes said processor to automatically:

pre-process the dataset to generate a pre-processed dataset via a data transformation of the dataset that makes the pre-processed dataset symmetric for retrospective analysis,

build a similarity matrix from the pre-processed dataset, and

output an optimum number of similarity clusters,

wherein the data transformation for retrospective analysis results from processing the dataset with a hazard function defined by;

G_t=α

*ln (G−

β

)−

α

*ln(α

), where parameter α

=T_c−

β

, and parameter β

=D_r−

1, in which T_c is a center of a transformed space, D_r is a minimum defined glucose level, G_t is the transformed data of the blood glucose concentration measurements provided in the dataset, and G is original glucose level values of the blood glucose concentration measurements provided in the dataset and measured in millimoles per liter.

17. The non-transitory computer-readable medium of claim 16, wherein the at least one time window dataset from a patient is a continuous glucose monitor (CGM) profile, and wherein said continuous glucose monitor (CGM) profile comprises raw data, transformed data, raw data associated with related data tags, transformed data associated with related data tags, or combinations thereof.

18. A method for identifying day(s) where a diabetes control therapy was inadequate for a patient using a monitoring system comprising a display device, a physiological data input device and a processor, the method comprising:

receiving automatically from the physiological data input device a plurality of blood glucose concentration measurements of the patient within a time window to generate at least one time window dataset of collected unsupervised daily monitoring profiles; and

executing from a memory a stored an unsupervised daily monitoring profile clustering algorithm and causing the processor automatically to;

pre-process the dataset to generate a pre-processed dataset via a data transformation of the dataset that makes the pre-processed dataset symmetric for retrospective analysis,

build a similarity matrix from the pre-processed dataset, and

output on the display an optimum number of similarity clusters found by the processor from the similarity matrix,

wherein the data transformation for retrospective analysis results from processing the dataset with a hazard function defined by;

G_t=α

*ln(G−

β

)−

α

*ln(α

), where parameter α

=T_c−

β

, and parameter β

=D_r−

1, in which T_c is a center of a transformed space, D_r is a minimum defined glucose level, G_t is the transformed data of the blood glucose concentration measurements provided in the dataset, and G is original glucose level values of the blood glucose concentration measurements provided in the dataset and measured in millimoles per liter.