Salman Haleem

Artificial Intelligence and Machine Learning for Digital Medicine and Health

With the advent of modern technology and the use of portable devices and wearables, we are able to acquire real-time status of physiological and physical activity. This may include vital signs (e.g., heart rate, blood pressure), calories burnt, daily steps count etc. The analysis and modelling of these kind of data can be used for human activity analysis, health trajectory analysis and non-invasive alternate to detect chronic events for improved quality of life. However, the fundamental understanding these kind of data has to be improved in order to address challenges associated with modelling real-time signals leveraging prediction of health trajectory for improved global wellbeing. This talk is concerned with fundamental design and development of data driven models with application towards healthcare and clinical decision support. This workshop will present about different modelling techniques in this direction and allow to think for possible solutions based on specific research question. Besides, it will allow to implement pragmatic solutions with Explainable AI solutions and possibilities to be integrated into real-world smart devices.

 

David Nickson

Depression prediction and diagnosis using machine learning with Electronic Health Records: replication and development.

Background

Recent advances in machine learning, combined with the growing availability of digitized health records, offer new opportunities for improving early diagnosis of depression. A growing body of research shows that Electronic Health Records can be used to accurately predict cases of depression based on an individual’s primary care records. The successes of these studies are undeniable, but there is a growing concern that their results may not be replicable, which could cast doubt on their clinical usefulness.

Method

We set out to reproduce and replicate the work by Nichols et al. (2018), who trained predictive models of depression among young adults using Electronic Healthcare Records. Our approach consisted of three parts. First, we attempted to replicate the methodology used by the original authors, acquiring a comparable set of primary electronic health records, and reproducing their data processing and analysis. Second, we tested the models presented in the original paper on a non-overlapping subset of our data, thus providing an out of sample prediction for the predictive models. Third, we extended past work by considering several novel machine learning approaches in an attempt to improve the predictive accuracy achieved in the original work.

Results

Our results demonstrated that the work of Nichols et al. (2018) is reproducible and replicable. We demonstrated an improved performance over the original models (average AUC 0.88 vs 0.72) that was maintained for the out of sample predictions (average AUC 0.85). This may be due to changes in mental health awareness over time, to operational changes in, e.g., the Quality Outcomes Framework or National Institute for Health Care and Excellence guidelines or data collection methods. Although alternative predictive models did not improve performance over standard logistic regression, our results indicated that stepwise variable selection is problematic due to its inconsistency.

Conclusion

Based on our results, predictive models of this type can be robust and replicable. We identified challenges with the nature and variability of Electronic Health Records, the predictors of depression and factors impacting on generalizability. However, providing these can be overcome such models can be made auditable, interpretable, maintainable, and potentially clinically useful.

 

Tim Dong

Advancements in Ensemble Learning for Cardiovascular Risk Prediction

Objective
The introduction of new clinical risk scores (e.g. European System for Cardiac Operative Risk Evaluation (EuroSCORE) II) superseding original scores (e.g. EuroSCORE I) with different variable sets typically result in disparate datasets due to high levels of missingness for new score variables prior to time of adoption. Little is known about the use of ensemble learning to incorporate disparate data from legacy scores. We tested the hypothesised that Homogenenous and Heterogeneous Machine Learning (ML) ensembles will have better performance than ensembles of Dynamic Model Averaging (DMA) for combining knowledge from EuroSCORE I legacy data with EuroSCORE II data to predict cardiac surgery risk.
Method
Using the National Adult Cardiac Surgery Audit (NACSA) dataset, we trained 12 different base learner models, based on two different variable sets from either EuroSCORE I (LogES) or EuroScore II (ES II), partitioned by the time of score adoption (1996-2016 or 2012-2016) and evaluated on holdout set (2017-2019). These base learner models were ensembled using nine different combinations of 6 ML algorithms to produce homogeneous or heterogeneous ensembles. Performance were assessed using a consensus metric.
 
Results
Xgboost homogenous ensemble (HE) was the highest performing model (CEM 0.725) with AUC (0.8327; 95% Confidence Interval (CI) 0.8323-0.8329) followed by Random Forest HE (CEM 0.723; AUC 0.8325; 95%CI 0.8320-0.8326). Across different heterogenous ensembles, significantly better performance was obtained by combining siloed datasets across time (CEM 0.720) than building ensembles of either 1996-2011 (t-test adjusted, p = 1.67e-6) or 2012-2019 (t-test adjusted, p = 1.35e-193) datasets alone.
Conclusion
Both homogenous and heterogenous ML ensembles performed significantly better than Dynamic Model Averaging ensemble of Bayesian Update models. Time-dependent ensemble combination of variables, having differing qualities according to time of score adoption, enabled previously siloed data to be combined, leading to increased power, clinical interpretability of variables and usage of data.

 

 

Thomas Popham

AI in Healthcare - what does this mean for education?

Data analytics and machine learning technologies are set to grow within the healthcare sector, but how are Universities responding to this change from an educational point of view? This talk looks at the growing number of courses in healthcare analytics and analyses the trends. A short overview will also be given of the new Diagnostics, Data and Digital Health MSc course which will start at Warwick in October 2023 and is a multi-disciplinary course involving modules taught by Warwick Medical School, School of Engineering, WMG, Chemistry and IATL.

 

Melina Kunar

How to Improve Human Interaction with AI in Healthcare

Artificial Intelligence (AI) has the potential to improve healthcare in relation to medical screening. Computer Aided Detection (CAD), for example, is being used to flag suspicious areas in medical screening to help clinicians detect abnormalities. Despite the dramatic investment into and increased use of this technology, human readers (e.g., clinicians) are still needed to interact with AI and make the final decisions (at least for now). We have previously found that this interaction is beneficial when AI accurately highlights suspicious items, however, there is a cost where readers become over-reliant on the technology. In this case, readers are more likely to miss cancers when CAD fails to flag a suspicious area than when no CAD system is employed. Furthermore, readers show an increased proportion of false alarms with inaccurate CAD prompts. Along with advancements in AI there needs to be greater investigation of how best to present this technology to humans to keep its benefits, while mitigating costs of over-dependence. This talk will showcase research investigating optimal ways to present CAD to human readers to determine best outcomes in cancer detection (e.g., presentation modes, instruction sets, etc.). Our studies have shown that alongside investment in technology it is equally important to investigate the different methods of presenting AI to humans for optimal interaction.

 

Davide Piaggio

AI in Medicine - Considerations on regulatory aspects: enabler or barrier of progress?

The medical device sector is expanding at unprecedented rates, compared to that of pharma and biotechnologies. Artificial intelligence is pervading more and more every aspect of our life, including health and healthcare. Despite promising results and opportunities that the recent developments carry about in this sector, strict and improved regulatory frameworks are the necessary step to warrant safe and effective innovation. This talk will give an overview of how artificial-intelligence-based medical software is classed as a medical device, of the trends of AI in medicine and of the related risks and harms.

 

Robert Hollingworth

Research Impact Support at Warwick

If you want to see your research into AI and healthcare make a difference to patients, or other beneficiaries outside of academia, then the Warwick Impact Team can help. Research Impact Manager Rob Hollingworth will discuss the role of the Impact Team, the internal funding available to support translational research, and advise on how to make a successful impact funding application.

 

Josh Cartwright

Introduction to Warwick Innovations

Business Development Manager Josh Cartwright will provide an introduction to Warwick Innovations. The presentation includes how the team works with researchers to protect intellectual property so it can go on to see real-world use through licensing or spinout formation, and provides examples of the teams’ support for AI in Healthcare.