Development of data mining technique to detect drug adverse events (DAE) from health insurance claims
| Project/Area Number |
14570373
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Public health/Health science
|
|---|
| Research Institution | National Institute of Public Health |
|---|
Principal Investigator |
OKAMOTO Etsuji National Institute of Public Health, Department of Management Sciences, Section Chief, 経営科学部, 経営管理室長 (90247974)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
大坪 浩一 国立保健医療科学院, 技術評価部, 主任研究官 (80248559)
林 謙治 国立保健医療科学院, 次長 (00124404)
|
|---|
| Project Period (FY) |
2002 – 2005
|
| Project Status |
Completed (Fiscal Year 2005)
|
| Budget Amount *help |
¥3,200,000 (Direct Cost: ¥3,200,000)
Fiscal Year 2005: ¥700,000 (Direct Cost: ¥700,000)
Fiscal Year 2004: ¥700,000 (Direct Cost: ¥700,000)
Fiscal Year 2003: ¥700,000 (Direct Cost: ¥700,000)
Fiscal Year 2002: ¥1,100,000 (Direct Cost: ¥1,100,000)
|
|---|
| Keywords | health insurance claims / data mining / post market surveillance / binomial test / drug adverse event / Cartesian product / pharmacovigilance / patient safety / カイ二乗検走 / カイ二乗検定 / ニューラルネットワーク / プロペンシティスコア / 小柴胡湯 / 間質性肺炎 |
|---|
| Research Abstract |
Drug adverse events (DAEs) are hard to detect because they are mostly unexpected and rare in incidence. Spontaneous reporting from pharmaceutical companies and doctors/pharmacists may not be satisfactory because of it is subject to biases inherent in its very "spontaneous" nature. Health insurance claims are administrative data and differ from spontaneous reporting in that they have a well-defined population at risk because all office visits and prescriptions are captured. This means that claims data have denominator and hence it is possible to calculate rate. On the other hand, it has limitations : since claims capture all diagnoses and prescriptions, they inevitably contain much "noise". Data mining technique to overcome these limitations was developed and applied to a commercially collected health insurance claims database. The proposed method is 1) developing Cartesian products (combinations) of drugs and diagnoses using SQL language, 2) weeding out "noise" bye applying Poisson distribution excluding the Cartesian products whose frequencies are less than average plus SD^*1.96 and 3) testing the_probability of the dispensing date of drugs preceding the date of diagnoses by binomial distribution. If a drug is shown to have high preceding rate to the date of diagnoses of suspected DAEs, one can not exclude the possibility of the drug causing the diagnoses. This method was applied to a large claims dataset collected by Japan Medical Data Center (JMDC) and demonstrated a potential link between Terbinafene (antifungal drug) and liver damage. This side effect is already well documented in the publicly available spontaneous reporting data. The proposed method was demonstrated to be able to detect potential cause-effect relationship between drugs and diagnoses and is likely to be an effective data mining technique if a national database of computerized health insurance claims become operational in the near future.
|
Report
(5 results)
Research Products
(18 results)
