Tech Plus
Data mining and machine learning look set to revolutionise knowledge management in routine business processes[1]. Can they be used to assist with the identification and evaluation of new liability risks? The Robotic Toxicologist report presents an expert-based analysis of a liability-related work recently published by Allianz[2] et al.
The Allianz et al. report demonstrated some very appealing underlying capabilities. However, it is clear there are areas where targeted expert assistance could lead to improvements in toxicological insight, relevance and liability meaning-making. Insurers wishing to develop or acquire robotic tools could include such targeted expertise in their development and evaluation projects and if they are adopted, in the management of machine outputs.
The Robotic Toxicologist Report.
Could an automated search machine usefully identify injury outcomes associated with three of the chemicals in nail varnish? There were 15 overt “predictions” in the Allianz et al. report, so, if any single prediction is to be assumed to be more reliable than tossing a fair coin there should be at least 8 true positives and at most 7 false positives. A detailed analysis of how well they did is available[3] to UK liability insurers and reinsurers for £16k plus VAT. 32 pages. Subscribers[4] to the Radar service have already received The Robotic Toxicologist report, in confidence, as part of the service.
Related services
It seems likely that many insurers will be developing or accessing machine tools and data management techniques to help identify emerging liability risks. Such developments would benefit from:
Validation and Prioritisation:
A scenario based approach can be used to identify true positives, true negatives, false positives and false negatives. Machine predictions would be compared with expert assessment of the toxicology and mechanisms for liability exposures. Those predictions which seem reasonable would also have a better chance of being accepted in principle by insurance purchasers.
Machine Education:
Feedback from the expert assessment can be used to help the machine to self-optimise. The machine needs to ‘get a handle’ on the quality and relevance of science publications as well as the stated[5] findings. A useful machine would also account for the binary tests of liability exposure[6].
Quantification:
The likely exposure associated with a true positive “event not in data” (ENID) can be calculated using epidemiological methods. These methods have been developed for Radar subscribers and the same techniques can be applied quite generally. The method is transparent, adapted to sensitivity testing, suited to asking the “what if” questions and all data inputs are taken from verifiable quality assessed sources.
Although each ENID model is unique, the same principles have been applied to:
- night shift work as a latent cause of breast cancer, (EL/WC claims)
- e-cigarettes as a cause of inherited asthma, (Product liability, Pharmaceuticals)
- an occupational cause of heart disease, (EL/WC claims)
- an occupational cause of stroke, (EL/WC claims)
- processed meat as a cause of two kinds of cancer (Product liability)
- the effect of antimicrobial resistance on some cancer claims (Product liability (for the causal agent)/Med Neg)
- (UK) vibration white finger (EL claims)
These ENID models are available for sale to UK liability insurers and those with a specialist interest e.g. processed meat manufacturers or the manufacturers of e-cigarettes and e-liquids. Each can be readily adapted to other jurisdictions once suitable data has been found. Some are directly mapped onto SIC (NAICS) codes.
Contact:
Dr Andrew Auty
andrew@reliabilityoxford.co.uk
[1] https://www.linkedin.com/pulse/artificial-intelligence-liability-insurance-andrew-auty/
[2] https://www.agcs.allianz.com/assets/PDFs/risk%20bulletins/AGCS_Praedicat_Toxic_Trio_Risk_Bulletin.pdf
[3] For more on this see: http://www.reliabilityoxford.co.uk/robot-toxicologist-or-myth-machine/ . The Robotic Toxicologist report is provided in confidence. Enquiries from non-insurers will also be considered.
[4] http://www.reliabilityoxford.co.uk/radar/
[5] Science authors are free to select the statements they make. It is quite normal that one corroborative fact may be highlighted while ten which are opposed are not described in the text. Text-reader machines must be able to test whether the statements in the paper are representative and if not representative, does it matter?
[6] Proof of causation, breach, insurance cover etc.
