Machine Learning & Heart Attack Risk

Overview

  • traditional medical systems
    • treat illness when people are already sick
  • the reason of using machine learning
    • more accurate
      • 20% - 30% better performance (means better prediction)
    • “unbiased”
      • statistical vs prejudicial bias
    • fast
      • automated decision making system
      • but some decision making system may generate risks
    • cheap
      • predictive models work 24/7
      • can provide medical services all the time

Scorecard

  • features
    • easy to understand and fits with our common sense and intuition
    • it is “explainable” or “white-box”
  • scenario - predicting heart disease
  • predicting heart disease with scorecard
  • likelihood of heart attack in the next 5 years
  • decision rules & cutoffs
    • if this was a Covid-fatality risk calculator, who would you vaccinate first & why
    • if this was a default risk calculator, who would you deny credit to, who would you loan to
      • Fair Isaacs Corporation created FICO, a credit scoring system used in the US
    • consider age
    • yong or low-income families pay more to buy a house, with a higher mortage interest rate
  • FinTech - the disrupting force
    • different data sources
    • applying AI, machine learning and predictive analytics for credit scoring

Bank loans & credit in China

  • history
    • 1949 all FIs nationalized -> People’s Bank of China
    • 1980s Bank of China, Agricultural Bank of China, China Construction Bank, Industrial & Commercial Bank of China established for loans to SOEs
    • 2010 all listed on stock exchange, largest banks in the world by market cap, 1⁄2 of total assets in China
    • regulatory control of interest rates for insurance and mutual fund products
    • 20% of SMEs had access to credit, 60% GDP
    • private micro-enterprises of < 10 employees, no credit access
    • 2008 PBOC and CBRC microfinance defined
    • 2013 village and town 7839 microfinance companies with $131B in loan balance
  • CreditEase
    • poor talented students seeking vocational training at Ning Tang’s (the founder) school
    • problem - students can’t afford tuition
    • need - speedy credit approval for tuition loan
    • risk — default, not paying back
    • reward— better-paying jobs increase likelihood of payback
    • proxies — grades, attendance, teacher recommendations
    • credit rating system
  • P2P lending
    • the importance of credit-scoring
  • bank loans vs P2P lending
    • the source of money is different
    • the risk control of traditional bank is better than that of P2P platform
  • Dianrong & sesame credit
    • using alternative data - “data exhaust”
  • Ant Financial & sesame credit
    • consumption scenarios
    • creating real-time datastreams

AI-Enabled Insurance

Overview

  • insurance technology
    • is an important sector of the financial services industry
  • the reason of using InsurTech
    • more accurate
      • better predictions and better at spotting claims fraud via video, e.g. Lemonade 1 min. claim
    • “unbiased”
      • reduced agent favoritism or outright corruption, e.g. Ping An Medical Insurance/OneConnect
    • fast
      • speeds the application and ID verification & approval process and the customer experience, e.g. Ant Financial auto 6 seconds vs 6 days for claims reimbursement
    • cheap
      • more accessible medical care and emergency payment for rural and underserved areas, e.g. The Good Doctor telemedicine one-minute clinic

Lemonade’s AI bot is easy and fast

  • video description

  • top 5 data questions

    • home address
    • own or rent
    • roommates, fire alarm, burglar alarm
    • birthdate
    • active renter’s insurance

Ethical Considerations

Assessing the risk of an automated decision making system

  • beneficiary
    • who is going to gain from the decision being made
    • high risk from high benefits for the decision maker
    • low risk from low benefits for the individual
    • because of a unbalanced situation
  • data immutability
    • if decisions are made on the basis of characteristics people are born with such as age and ethnic origin
    • this is higher risk and more controversial than lifestyle choices such as the music people like or what they watch on TV
  • impact
    • what effect is a decision going to have
    • a life or death decision about cancer treatment is much more important than whether or not to send someone a 10% coupon for frozen pizza

Approaches of using predictive models for “high risk” decisions

  • try to identify at risk groups
    • separate score distribution reports can be compared to see if certain groups will be adversely treated
  • a more fair treatment
    • for lower than average scoring groups, constraints and over-ride rules as well as different cut-offs can be set
  • continue to monitor the situation once the decision making system goes live
    • review score profile and decision rules for key groups
    • fine tune cut-offs, constraints, and over-rides

InsurTech

  • Axa - $10k large-loss claims

    • video description (until 9'20'')

  • ethical considerations for Axa

    • what are the automated decision rules that will use the predictions of the large-claim predictive model
    • who might be at-risk groups
    • what constraints, over-ride, cut-offs should be set for risk groups
    • how to monitor
    • how to protect reputation if the large-claim predictive model is used to “screen out” undesirable drivers

  • different legislative regimes

    • US utilitarian approach
      • personal data can be harvested for business goals
      • if there is a problem using a specific type of data, unacceptable bias against a specific group, legislation is enacted to address that concern
    • European Union rights based approach
      • personal ownership of data, companies have no right to hold or use personal data unless the individual gives permission
      • GDPR (General Data Protection Regulations) for EU citizens and transactions, “right to erase”

Additional Reading