The Cost of Chronic Conditions
A key focus for care management and predictive analysis
With rising costs and limited budge...
Predict Year 2 costs / risk for a general diabetes group
A group of members with diabetes were selected – including a wide...
A key focus for care management and predictive analysis
The Cost of Chronic Conditions
Data sources
One year of standard m...
Initial results: PreVista neural network
An initial ‘test’ run with the neural network produced R2 accuracy
of only .32 (3...
Training run 2: neural network with GA tuning
The PreVista genetic algorithm was added, to identify and weight key factors...
Training run 3: neural network with additional GA rounds
Genetic algorithms are process intensive. Each PreVista run with ...
Additional findings
The combination of genetic, fuzzy and neural network technology often uncovers interesting
relationshi...
For more information please contact:
Improved accuracy with PreVista
Mark Hays
CEO
Phone: 508.661.9733
Email: MarkHays@Eas...
of 8

PreVista - diabetes focus - v3b - 12.18.2014

Published on: Mar 4, 2016
Source: www.slideshare.net


Transcripts - PreVista - diabetes focus - v3b - 12.18.2014

  • 1. The Cost of Chronic Conditions A key focus for care management and predictive analysis With rising costs and limited budgets, we need to identify causes and solutions: “Which members should we target for intervention?” and “How can we reduce readmissions?” With Medicare, for example, ~60% of total costs are driven by 10% of the population and cost growth is fueled by ten chronic conditions as shown in the graphs to the right. This familiar graph highlights the key issue: payers and providers need to identify at-risk members and cost drivers, manage risk, target interventions and improve care. All of these changes require improved ‘intelligence’ to guide the healthcare process. PreVista™ can target each chronic condition – and combinations of chronic conditions – to identify the most at-risk members and the treatments that are most effective. The following PreVista example focuses on diabetes, one of the most complex challenges -- with a wide age range of members and co-morbid conditions. Page 1
  • 2. Predict Year 2 costs / risk for a general diabetes group A group of members with diabetes were selected – including a wide range of ages and co-morbid conditions. This was a chaotic data set with many wide-ranging variables that made prediction particularly difficult: 3,937 members Age range: 1 to 91 Outpatient visits: 1 to 325 Number of co-morbid conditions: 0 to 32 Year 1 costs: $3 to $421,915 Change, Year 1 to Year 2: -$285,883 to +$313,423 Page 2 A key focus for care management and predictive analysis The Cost of Chronic Conditions
  • 3. A key focus for care management and predictive analysis The Cost of Chronic Conditions Data sources One year of standard medical and pharma claims were used as inputs, along with a few basic fields from the client’s care and disease management system. (Shown to the right) Lab and EMR data was not yet available for these members. This more current and detailed data would be very helpful with PreVista. Age Gender Frailty flag Medical claims cost - prior year Pharmacy cost - prior year Active ingredients - count Rx gaps - count Major procedures - count Inpatient hospitalizations - count ED visits - count Outpatient visits - count Management visits - count Providers - count Specialists – count Total co-morbid conditions – count Hospital dominant conditions - count Nursing services Dialysis services Immuno suppression treatment Chronic obstructive pulmonary disease Asthma Ischemic heart disease Congestive heart failure Renal failure Hypertension Lipid metabolism disorders Osteoporosis Low back pain Rheumatoid arthritis Hypothyroidism Age related macular degeneration Pregnancy without delivery Seizure disorders Glaucoma Depression Bipolar disorder Page 3
  • 4. Initial results: PreVista neural network An initial ‘test’ run with the neural network produced R2 accuracy of only .32 (32%). This would be a good result for standard ‘predictive modeling’ and ‘risk score’ products. In the graph below, blue = actual costs and red = predicted: Page 4 Technical note: The cost range is limited to $200,000 due to the elimination of 90 ‘outlier’ records which would skew the training process. An example of PreVista™ results with diabetes The Cost of Chronic Conditions
  • 5. Training run 2: neural network with GA tuning The PreVista genetic algorithm was added, to identify and weight key factors in the data. These results were loaded into the neural network. R2 accuracy jumped to .56 (56%). The green section of the graph shows the delta between predicted and actual results: Page 5 Technical note: The cost range is limited again to $200,000 due to the elimination of 90 ‘outlier’ records which would skew the training process. An example of PreVista results with diabetes The Cost of Chronic Conditions
  • 6. Training run 3: neural network with additional GA rounds Genetic algorithms are process intensive. Each PreVista run with the diabetes data set included 100 rounds and required 4 days of computation time on an R&D server. 300 rounds were added and fed to the neural network. R2 accuracy increased to .70 (70%). Page 6 Technical notes: • The cost range increased in this graph to $350,000 after the outliers were included, to test the trained model against the entire data set. • The sample population used for this analysis was diverse in age and acuity. The resulting predictive model is not ‘over trained’ to a narrow data set, and can be re-used without additional training. An example of PreVista results with diabetes The Cost of Chronic Conditions
  • 7. Additional findings The combination of genetic, fuzzy and neural network technology often uncovers interesting relationships in the data. In this analysis of diabetes, for example, we found: Although 90 outliers were excluded while the PreVista model was trained, analysis was particularly accurate with this set of records – near 100%. Review of the data showed that the cases were linked to specific co-morbid conditions, e.g. renal failure. This is an example where a specific model would be trained for members who match the high-risk subset within the diabetic population. These members would be a high priority for intensive care / disease management. Members with Year 1 costs of ~$40K typically showed a reduction in Year 2 of 50%. Members with costs between $1K and $3K often saw an sharp increase in Year 2 of ~10X. These members are prime targets for improved care / disease management. Page 7 An example of PreVista results with diabetes The Cost of Chronic Conditions
  • 8. For more information please contact: Improved accuracy with PreVista Mark Hays CEO Phone: 508.661.9733 Email: MarkHays@EastBayLabs.org “East Bay Labs”, the East Bay Labs logo and “PreVista” are trademarks of East Bay Labs. This document and the contents contained herein are copyright 2012, 2013, 2014, 2015 East Bay Labs, with the exception of images, product names and references belonging to other companies. East Bay Labs is a social benefit corporation dedicated to improving the quality and effectiveness of healthcare. Page 8

Related Documents