Predictive Diagnosis

Heart Attacks

  • Heart attack is a common complication of coronary heart disease resulting from the interruption of blood supply to part of the heart

  • 2012 report from the American Heart Association estimates about 715,000 Americans have a heart attack every year
    • Every 20 seconds, a person has a heart attack in the US
    • Nearly half occur without prior warning signs
    • 250,000 Americans die of Sudden Cardiac Death yearly
  • Well - known symptoms
    • Chest pain, shortness of breath, upper body pain, nausea
  • Nature of heart attacks make it hard to predict, prevent and even diagnose
    • 25% of heart attacks are silent
    • 47% of sudden cardiac deaths occur outside hospitals, suggesting many do not act on early warning signs
    • 27% of respondents to a 2005 survey recognized the symptoms and called 911 for help

Analytics Help Monitoring

  • Understanding the clinical characteristics of patients in whom heart attack was missed is key

  • Need for an increased understanding of patterns in a patient’s diagnostic history that link to a heart attack

  • Predicting whether a patient is at risk of a heart attack helps monitoring and calls for action

  • Analytics help understand patterns of heart attacks and provides good predictions

Claims Data

  • Claims data offers an expansive view of a patient’s health history
    • Demographics, medical history and medications
    • Offers insights regarding a patient’s risk
    • May reveal indicative signals and patterns

  • We will use health insurance claims filed for about 7,000 members from January 200 - November 2007

  • Concentrated on members with the following attributes
    • At least 5 claims with coronary artery disease diagnosis
    • At least 5 claims with hypertension diagnostic codes
    • At least 100 total medical claims
    • At least 5 pharmacy claims
    • Data from at least 5 years

  • Yields patients with a high risk of heart attack and a reasonably rich history and continuous coverage

Data Aggregation

  • The resulting dataset includes about 20 million health insurance entries including individual medical and pharmaceutical records

  • Diagnosis, procedure, and drug codes in the dataset comprise tens of thousands of attributes

  • Codes were aggregated into groups
    • 218 diagnosis groups, 180 procedure groups, 538 drug groups
    • 46 diagnosis groups were considered by clinicians as possible risk factors for heart attacks

Diagnostic History

  • We then compress medical records to obtain a chronological representation of a patient’s diagnostic profile
    • Cost and number of medical claims and hospital visits by diagnosis
  • Observations split into 21 periods, each 90 days in length
    • Examined 9 months of diagnostic history leading up to heart attacks/no heart attack event
    • Align data to make observations date-independent while preserving the order of events
      • 3 months ~ 0 - 3 months before heart attack
      • 6 months ~ 3 - 6 months before heart attack
      • 9 months ~ 6 - 9 months before heart attack

Target Variable

  • Target prediction is the first occurrence of a heart attack
    • Diagnosis on medical claim
    • Visit to emergency room followed by hospitalization
    • Binary Yes/No

Dataset Compilation

Cost Bucket Partitioning

  • Cost is a good summary of a person’s overall health

  • Divide population into similar smaller groups
    • Low risk, average risk, high risk
  • Build models for each group

Predicting Heart Attacks (Random Forest)

  • Predicting whether a patient has a heart attack for each of the cost buckets using the random forest algorithm

Incorporating Clustering

  • Patients in each bucket may have different characteristics

Clustering Cost Buckets

  • Two clustering algorithms were used for the analysis as an alternative to hierarchical clustering
    • Spectral Clustering
    • k-means clustering

k - Means Clustering

Practical Considerations

  • The number of k can be selected from previous knowledge or experimenting

  • Can strategically select initial partition of points into clusters if you have some knowledge of the data

  • Can run algorithm several times with different random starting points

Random Forest with Clustering

Predicting Heart Attacks

  • Perform clustering on each bucket using k = 10 clusters

  • Average prediction rate for each cost bucket

Impact of Clustering

  • Clustering members with each cost bucket yielded better predictions of heart attacks within clusters

  • Grouping patients in clusters exhibit temporal diagnostic patterns within 9 months of a heart attack

  • These patterns can be incorporated in the diagnostic rules for heart attacks

  • Greater research interns in using analytics for early heart failure detection through pattern recognition