Rapid mass spectrometry coupled with machine learning to predict clinically relevant variables in cardiovascular disease

Thesis

Cardiovascular diseases create a large burden on global healthcare systems. Treating patients with cardiovascular diseases is challenging, as not all patients respond well to the same treatments. Identifying patients who need more intensive treatments, surgery, or monitoring is important in order to enable more personalised healthcare plans. Personalised healthcare, or tailoring medical treatment plans to the specific needs of a patient, has the potential to improve the lives of many patients, with the additional benefit of reducing the burden on healthcare systems through reducing readmission rates. Tools to allow this sort of care need to be developed.

The atmospheric solids analysis probe mass spectrometer (ASAP-MS) has been investigated for its potential for use as a clinical tool for predicting patient outcomes based on biological sample measurements. We have investigated this in the context of two cardiovascular diseases: ST elevated myocardial infarctions (STEMI), and abdominal aortic aneurysms (AAA).

We have developed an optimised protocol for the measurement of human blood plasma and tissue samples with ASAP-MS, discussed in Chapter 2. Using this optimised method, we can measure a metabolic fingerprint of biological samples. These fingerprints are complex to interpret, but can be mapped to clinical measurements of interest using machine learning methods, which are good at pattern recognition within complex data.

Using ASAP-MS measurements of blood plasma samples taken soon after STEMI patients first presented at the hospital, we were able to predict prognostically relevant information using machine learning methods. Highlights include the prediction of patient mortality with over 85% accuracy, and microvascular obstruction with over 83% accuracy. In the AAA study, patient groups with AAAs of different stages of development were classified with extremely high accuracy, at over 95%. AAA size was predicted with high accuracy at 93%, and indicators related to AAA growth were also found. Using both feature reduction methods and symbolic regression analysis, we have been able to start to identify the underlying relationships between specific metabolite peaks and clinical outcomes.

In the final chapters of this thesis, other aspects of the ASAP-MS method were investigated, including a comparison of the commercial instrumentation available and assessment of the impact of normalisation methods on ASAP-MS data.

Through this investigation, we have provided a proof-of-concept study showing that combining ASAP-MS and machine learning techniques can provide clinically useful information for the rapid prediction of disease outcomes. This provides further evidence to support the growing use of techniques based on ambient ionisation mass spectrometry in clinical environments.

Authors: Eardley-Brunt, ASJ

• DOI: 10.5287/ora-eoo9xpjpo
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: ASAP mass spectrometry; clinical mass spectrometry; clinical machine learning; ambient ionisation mass spectrometry; normalisation
• Subjects: Clinical data analysis; Machine learning; Cardiovascular disease; Chemistry; Analytical chemistry