Causal Diagnostics
Despite the many advances in medical technology, practical clinical diagnosis is still largely based on statistical symptom pattern-matching. The idea that a cluster of symptoms can lead to the correct diagnosis is often an oversimplification. Not only are symptoms not binary (either one has symptom or does not), but complex diseases are multifactorial, and can display a wide number of dissimilar effects and confounding signals. To better deal with probable root causes, OIA leverages our years of research in model-driven causal discovery at some of the best universities (Oxford, Cambridge, Karolinska, Imperial) to develop the area of causal diagnostics, a cutting-edge and responsible, model-based approach to medicine that differs from simplistic AI approaches based on statistical classification that are generally unable to provide generative models or explanations for their arbitrary choices.
At OIA we have created and continue creating a family of novel and interdependent remote and mobile solutions for rapid and affordable generation of personal longitudinal data using sophisticated neuro-symbolic machine learning. We repurpose off-the-shelf electronics and augment them with this responsible AI. Our mission is to build the first portable and personal general-purpose medical device to help with patient monitoring and super-early diagnosis.
|
Longitudinal, non-invasive, (quasi) real-time and cost-effective health monitoring is hugely important if we are to improve public health and deliver on the promise of personalised medicine that can enable super-early, accurate diagnosis.
Causal diagnostics can revolutionise medicine by refocusing on first principles and causal generating mechanisms to redesign treatment and drugs to go beyond merely offering symptom relief. The training and inference engines behind causal diagnosis are based on what is identified as neuro-symbolic computation, a combination of the best AI tools with cutting-edge expert systems that rely on both mind and machine to fight disease. 
|
|
|
|
We call our neuro-symbolic computation approach a type of responsible Artificial Intelligence (AI) because it goes beyond current black-box approaches in AI in order to reach a deeper understanding of critical issues and eventually unveil root causes in various areas of medicine.
|
|
Responsible AI
|
While techniques such as machine & deep learning (e.g. deep neural networks) are very powerful in tasks related to classification, they are only so under very controlled conditions, and are ill-equipped to deal with basic cognitive functions that we as humans take for granted, including basic logical inference and abstraction for model generation and hypothesis testing.
The combination of symbolic computation and statistical machine learning is thus a a powerful one for tackling complex problems, and a way forward to achieve better and more responsible A.I. This is because it offers interpretable models that correspond to observations and can be tested rather than encoded in obscure black-box explanations. This is what OIA has been successfully pioneering, developing and implementing for several years, translating cutting-edge science to real-world human challenges. |
To illustrate the limitations of statistical AI, no amount of big data or statistical massaging (no matter how 'sophisticated') can make a traditional deep neural network perform a simple arithmetic operation if it has not seen it before because it has no concept of number or symbol and will treat the whole term as an object.
|
|
|
Personalised Healthcare
|
OIA has developed an innovative risk assessment score used in our remote blood testing and analytics solution, Algocyte. The measures were tested against 100 thousand cases, demonstrating that they pick up signals to separate healthy from unhealthy groups even from noisy sources (the NHANES database from the CDC) . At the centre (Score) is the numerical score that goes from 0 (healthy) to 10 (unhealthy), quantifying how removed a person's individual values are from reference population values. On the right (NCCIS) is the non-linear colour-coded score that further separates the healthy from unhealthy groups. The figure at left shows the left shows how the scores are displayed and utilised by Algocyte.
|
|
The score adapts and learns over time--with incremental precision--each patient's own baseline, allowing the application to flag for deviations smaller than the traditional reference population values, and thereby allowing a level of personalisation and sophistication not available before in the analysis of blood test results. On the left we see how the adaptive score follows changes captured by the non-adaptive (normal) score but--more importantly--how it adapts (figure at right) over time, resetting new normal v. abnormal reference intervals adapted to each patient.
|
|
These scores are at the centre of Algocyte's machine learning system for automating tasks such as patient triaging, and patient profile flagging.
|
|
|
Selection of publications
We have shown how our algorithms can deal with complex biological systems even in the face of sparse data by using model-based approaches combined with sophisticated techniques for extracting meaning from data.
|
|
|
|
World media coverage of the research methods behind our causal diagnostics approach
|
