Professional Documents
Culture Documents
Takahide Shinge
Yuji Kamata
Takanobu Osaki
Hirofumi Kondo
50
Figure 1 — Joint Research with Dai-ichi Life on Use of Medical Big Data
The development of hospitalization risk prediction models and other such technology, and its deployment in the insurance industry, is being investigated
by fusing expertise from both Dai-ichi Life and Hitachi.
Aims
Dai-ichi Life Differentiate from competitors by making
more sophisticated use of medical big data in
operations Medical knowledge
Challenges Insurance
Timely training of personal with expertise in
knowledge
Collaborative
both the insurance business and analytics creation projects
Hospitalization
risk prediction
Aims model,
Hitachi etc.
Establishment of services for insurance companies
that provide technology for analysis of medical costs
or other healthcare data Healthcare
cost prediction
Challenges
Identify actual needs of insurance companies and
build experience
background that Hitachi and Dai-ichi Life embarked Dai-ichi Life started by looking for ways in which it
on this joint research as partners with common aims could contribute to making acceptance criteria more
and facing common challenges (see Figure 1). fine-grained.
2. 1 2. 2
Insurance Acceptance Criteria Development of Prediction Model for Length of
Standard practice at life insurance companies is to Stay
base underwriting decisions about whether or not to The joint research involved the creation of a model
accept insurance applications on factors such as the that uses information on applicant health at the time
information provided by the applicant and the results of application to predict the length of future stays due
of a health check (a process called “risk selection”). to lifestyle disease. The model was developed by utiliz-
If an insurance company accepts large numbers of ing Hitachi’s analytics know-how and Dai-ichi Life’s
applicants with pre-existing conditions or with a high expertise in insurance medicine and medical under-
risk of dying in the near future, for example, the num- writing to analyze medical information from insur-
ber of claims will significantly exceed expectations, ance applications (included information provided by
undermining the viability of the insurance system. the applicant and that obtained from checkups) and
This makes it very important for insurance companies claim payment data built up by Dai-ichi Life over its
that they get risk selection right. many years in the health insurance business. While
However, the requirements of the insurance “insurance risk events” take a variety of forms (hospi-
industry include designing acceptance criteria that talization, surgery, illness, and so on), acquiring the
are more fine-grained than in the past so that more ability to predict risk in terms of the length of stay
people can take out insurance, both to fulfill public provides numerous insights into the optimization
expectations for private insurance and for reasons that of acceptance criteria in the risk selection process.
include changes in demographics and other aspects Accordingly, model development focused on hospi-
of the business environment in which the compa- talization due to lifestyle disease as a way to study the
nies operate. Accordingly, the joint research with criteria to use for such conditions.
HbA1c
BMI
Pre-existing
condition A
Disclosure
Analysis
dataset
Clinical
information Machine learning
Statistical analysis
Applicants seeking to take out life insurance typi- machine learning to identify which factors within this
cally have to disclose certain information about them- large volume of data contribute to lifestyle disease,
selves, including considerable medical information. combining it with a statistical analysis to incorporate
Moreover, because insurance policies have multi-year the clinical knowledge of physicians (see Figure 2).
terms, it is possible to acquire information on any When the resulting model was tested using trial
medical events that occur during this time after the data, it predicted length of stay with an accuracy of
policy is granted. Accordingly, Hitachi drew on its within 5%.
modelling expertise and past experience with the
analysis of clinical and medical claims data to use this 2. 3
information as the basis for developing a model that Use of Prediction Model for Length of Stay to
predicts the length of future stays based on informa- Loosen Acceptance Criteria
tion on an applicant’s health at the time of application. A study was conducted into using the model for pre-
The information dealt with by the health check dicted length of stay as a way to loosen the underwrit-
given to applicants can be broadly divided into general ing acceptance criteria. The approach adopted by the
clinical information and disclosure information spe- joint research was to assess the suitability of current
cific to private health insurance. The clinical informa- acceptance criteria by looking at the predicted length
tion includes various numerical scores such as those of stay for those people with some sort of abnormal-
obtained from blood tests, and the disclosure infor- ity in their health (such as the presence of conditions
mation includes large amounts of information such associated with high blood pressure, or abnormal
as the applicant’s medical history. The challenge for laboratory results for liver function) in comparison
the prediction of future hospitalization is to identify to those in normal health (see Figure 3).
which information in all of this is genuinely relevant As an example, the graph in Figure 3 shows the pre-
to ill health. Accordingly, the different types of data dicted length of stay for a person with a pre-existing
were first converted into forms that facilitate analysis condition A and a person in normal health plotted
to create a dataset of characteristic data expressed in against a particular laboratory result X [such as body
terms of both continuous (real-numbered) and binary mass index (BMI)]. The difference between the two
variables. As lifestyle disease was the subject of the indicates the expected value for length of stay.
study, the next step was to identify the risk factors Common practice at Dai-ichi Life has been to
that contribute to hospitalization due to lifestyle dis- impose conditions on anyone with the pre-existing
ease and to develop a model. This was done by using condition A for whom this particular laboratory
52
FEATURED ARTICLES
Person with
pre-existing condition A No large difference
Person without
pre-existing condition A
Under past criteria: Under past criteria:
Able to be assessed as low risk Assessed as high risk
15 20 25 30 35 40
Laboratory result X
result X exceeded a certain threshold, such as declin- hospitalization risk models to the insurance business
ing their insurance or raising the premium. However, that it has acquired through its involvement in the
the model indicates that the increase in hospitalization joint research with Dai-ichi Life.
risk with higher values of the laboratory result X is not While this joint research used data held by that
significantly greater for people with the pre-existing company, the challenge when developing general-
condition A than for those without. Accordingly, purpose risk models is to obtain the input medical
Dai-ichi Life revised certain of its previous criteria data. However, the increasing availability of medical
that took the form of “treat people with pre-existing big data in recent times, including health insurance
condition A who exceed a particular threshold for data being made commercially available, means that
laboratory result X as being high-risk” in those cases it has become possible to build hospitalization risk
in which previous proposals for the loosening of cri- models like that described above without utilizing
teria were backed up by the model result. As a result life insurance data. Risk models based on this health
of these criteria revisions, Dai-ichi Life expects to be insurance data can be used to predict the risk of future
able to accept around 2,000 applications annually that hospitalization based on information for individual
would previously have been rejected. policy holders (such as their being a male aged 45
with a BMI of 26). Whereas these numbers might
3. Establishment of Risk have resulted in an application being declined under
Simulation Service past underwriting criteria, the suitability of these cri-
teria are called into doubt if the model predicts only
This section describes the establishment and future a small increase in risk relative to a healthy person.
plans for a risk simulation service for insurance com- While the extent to which underwriting criteria (risk
panies that includes use of the hospitalization risk score under numerical rating system) should actually
model. be modified is a matter for the insurance companies
using the service to decide for themselves, the service
3. 1 can at least indicate where past underwriting criteria
Hospitalization Risk Simulation Service over- or under-estimate risks in the risk selection
Hitachi is planning to establish a risk simulation ser- process for individual applications (state of health)
vice that draws on the expertise in the application of (see Figure 4).
Conventional
underwriting Example criteria used by different life insurance
(numerical rating companies (product-specific conditions, etc.)
system) Product Conditions Assessment
Applicant information
AA policy Has pre-existing condition A or B Decline
Has pre-existing condition C and
Age: 45 BB policy
total risk score of 150 or more
Decline
Gender: Male ……
BMI: 26
HbA1c: 6.2
Blood pressure: 135/100
Liver function: ** Output Application
..... Total score +200 declined
Risk simulator
Output
Hospitalization risk
Underwriting department relative to healthy person:
(person responsible for 1.2 times Revise risk score or factors
revising criteria)
considered for applications that have
a small increase in hospitalization
risk relative to a healthy person
3. 2
Prospects for Risk Simulation Service
While this service is intended for use in underwrit- 4. Conclusions
ing by life insurance companies in Japan, it also has
potential applicavtions outside the insurance industry. This article has described insurance industry innova-
Given the increasing interest being shown by busi- tions that utilize expertise in the analysis of medical
nesses in staff health management, examples of such big data for insurance companies that Hitachi has
applications include use of the service to assess the built up over many years, together with examples of
risk of employee non-attendance, or its use as a means collaborative creation.
of assessing employee health risks in the services It is anticipated that the insurance and healthcare
provided by some banks that include the level of a industries will, in the future, put even more effort
company’s health management as a factor in setting into the development of new services that utilize not
finance interest rates. only medical big data but also data from the Internet
Looking outside Japan, Hitachi would also like to of Things (IoT) and personal health records (PHRs).
utilize the service in the development of new insur- As part of this, Hitachi will continue working on the
ance services through collaborations with some of the development of risk analysis technology.
steady stream of new startup businesses not beholden
to traditional risk underwriting methods that are
emerging in the US insurance market especially, or
in Southeast Asia where private life insurance is expe-
riencing rapid growth.
54
FEATURED ARTICLES
Acknowledgments Authors
Takahide Shinge
Hitachi received considerable advice and assistance Financial Information Systems Business Unit 4,
from The Dai-ichi Life Insurance Company, Limited Financial Information Systems 2nd Division, Hitachi,
Ltd. Current work and research: Development of
and other parties involved in the work described in new businesses in collaboration with an insurance
this article. The authors would like to take this oppor- company.