Cyber risk and Hazard risk.
Challenge – Addressing a partially insurance or currently uninsurable risk.
Context: Every two years, Aon releases the Risk Management survey whereby Aon asked over 2,300 risk managers and c-suite professionals from 60 countries/territories and 16 industries about their key risks and how they manage and mitigate them. A detailed report is produced and the top 60 emerging risks are announced – some of which are either only partially insurable or uninsurable. (One of key the reasons being it is challenging to price the risk).
View the 2021 GRMS survey & results: Executive Summary - 2021 Global Risk Management Survey (aon.com)
Opportunity: Select one of the partially insurable or uninsurable risks listed and build a model utilizing open source data (or combination of Aon data) to better understand or ‘price’ the risk so it can be developed in the future into an insurable product.
Parkison’s disease is a movement disorder manifesting itself as hypokinesia, slowness of movements. Parkinson’s patients are commonly assessed by MDS-UPDRS rating scale that consists of several exercises. This assessment can be captured by video and analysed using computer vision techniques. Deep learning methods allow the detection of keypoints such as fingertips, which in turn can be converted to a clinically relevant signal. One example of such a series is a Finger tapping signal, where the patient repeatedly taps the index finger and the thumb generating a periodic signal. Frequency, amplitude, or abnormalities in such movements are indicative of the disease.
For this project we propose analysing these series for the detection of clinically relevant information:
1) Develop robust and automated peak detection algorithm for the periodic signals.
2) Develop a hesitation detection algorithm (anomaly when the patient slows or interrupts the movement for a short period).
As an extension of the project we have two follow-up tasks if time permits:
3) Clustering patients into clinically relevant subtypes using the clinical UPDRS scores. This analysis could be repeated for computer vision models developed by us for UPDRS to see if similar subgroups arise.
4) Supervised classification of the series according to the UPDRS score.