Samuel W. Channon-Wells (United Kingdom)

Imperial College London Department of Infectious Disease
I'm a Paediatric trainee from Oxford, UK. My main interests are in Paediatric Infectious Disease diagnostics, and antimicrobial use in children. I am particularly interested in using big data to solve key challenges in Paediatric research, focusing predominantly on transcriptomics of infectious diseases, building on my previous training as a Mathematician. I'm currently working as a clinical research fellow on the very exciting DIAMONDS project at Imperial College London, which is aimed at developing new transcriptomic-derived diagnostics for Paediatric Infectious and Inflammatory disorders. Through this role I am excited to now also be working on the BATS trial, leading the latest updated analysis of this exciting clinical trial into treatments for MIS-C.

Author Of 2 Presentations

Debate 1: Should The Neonatal Early Onset Sepsis Calculator be Used Outside the USA (yet)? Yes or No

Date
Thu, 12.05.2022
Session Time
08:00 - 09:30
Session Type
Special Session
Room
BANQUETING HALL
Lecture Time
08:27 - 08:52

VALIDATION OF TRANSCRIPTOMIC SIGNATURES FOR FEBRILE CHILDREN USING NANOSTRING TECHNOLOGY AND EXPLORATION OF MULTI-CLASS PREDICTION MODELS

Date
Wed, 11.05.2022
Session Time
15:40 - 17:15
Session Type
Parallel Symposium
Room
DIMITRIS MITROPOULOS HALL
Lecture Time
16:53 - 17:01

Abstract

Backgrounds:

Many host transcript signatures for paediatric inflammatory and infectious diseases are in development, but require validation in independent cohorts; their translation to clinically useful test platforms lags behind discovery. We used NanoString technology to efficiently validate multiple signatures in parallel and explore the potential for more sophisticated multi-class classification models.

Methods

We validated five transcriptomic diagnostic signatures using prospectively recruited patients from multiple paediatric cohorts. Final phenotypes were assigned using pre-agreed definitions after review of clinical and laboratory data. We quantified 69 transcripts on a custom NanoString nCounter cartridge, normalising expression values using reference genes. Signature performance was assessed using Area Under ROC Curve (AUC) statistics. We explored two approaches to multiclassification diagnostics to develop proof-of-concept methods: a mixed test combining four independent one-vs-all models, and a multinomial model.

Results:

Our cohort of 92 paediatric patients included 23 definite bacterial and 20 definite viral infections, 15 Kawasaki disease, 18 with tuberculosis and 16 healthy controls. The signatures achieved AUCs above 0.82 (Table 1), with confidence intervals overlapping those of the respective discovery studies. However, performance declined in all signatures when tasked with differentiating additional groups. For example, the single-transcript BATF2 had AUC of 0.910 differentiating TB from healthy individuals, reducing to 0.745 when differentiating TB from other febrile diseases. In comparison, the multinomial approach identified a 24-transcript model that correctly classified all 76 non-control patients (0% in-sample error), outperforming the mixed-model (19 transcripts, 19.8% in-sample error).

table1.jpg

Conclusions/Learning Points:

The cross-platform, out-of-sample findings validated 5 signatures, but discriminatory power was reduced in patients drawn from outside their remit. An exploratory 24-transcript model had improved accuracy across all diagnostic groups, demonstrating in principle the utility for one-step multi-class diagnosis in patients with broad diagnostic uncertainty.

Hide

Presenter of 2 Presentations

Debate 1: Should The Neonatal Early Onset Sepsis Calculator be Used Outside the USA (yet)? Yes or No

Date
Thu, 12.05.2022
Session Time
08:00 - 09:30
Session Type
Special Session
Room
BANQUETING HALL
Lecture Time
08:27 - 08:52

VALIDATION OF TRANSCRIPTOMIC SIGNATURES FOR FEBRILE CHILDREN USING NANOSTRING TECHNOLOGY AND EXPLORATION OF MULTI-CLASS PREDICTION MODELS

Date
Wed, 11.05.2022
Session Time
15:40 - 17:15
Session Type
Parallel Symposium
Room
DIMITRIS MITROPOULOS HALL
Lecture Time
16:53 - 17:01

Abstract

Backgrounds:

Many host transcript signatures for paediatric inflammatory and infectious diseases are in development, but require validation in independent cohorts; their translation to clinically useful test platforms lags behind discovery. We used NanoString technology to efficiently validate multiple signatures in parallel and explore the potential for more sophisticated multi-class classification models.

Methods

We validated five transcriptomic diagnostic signatures using prospectively recruited patients from multiple paediatric cohorts. Final phenotypes were assigned using pre-agreed definitions after review of clinical and laboratory data. We quantified 69 transcripts on a custom NanoString nCounter cartridge, normalising expression values using reference genes. Signature performance was assessed using Area Under ROC Curve (AUC) statistics. We explored two approaches to multiclassification diagnostics to develop proof-of-concept methods: a mixed test combining four independent one-vs-all models, and a multinomial model.

Results:

Our cohort of 92 paediatric patients included 23 definite bacterial and 20 definite viral infections, 15 Kawasaki disease, 18 with tuberculosis and 16 healthy controls. The signatures achieved AUCs above 0.82 (Table 1), with confidence intervals overlapping those of the respective discovery studies. However, performance declined in all signatures when tasked with differentiating additional groups. For example, the single-transcript BATF2 had AUC of 0.910 differentiating TB from healthy individuals, reducing to 0.745 when differentiating TB from other febrile diseases. In comparison, the multinomial approach identified a 24-transcript model that correctly classified all 76 non-control patients (0% in-sample error), outperforming the mixed-model (19 transcripts, 19.8% in-sample error).

table1.jpg

Conclusions/Learning Points:

The cross-platform, out-of-sample findings validated 5 signatures, but discriminatory power was reduced in patients drawn from outside their remit. An exploratory 24-transcript model had improved accuracy across all diagnostic groups, demonstrating in principle the utility for one-step multi-class diagnosis in patients with broad diagnostic uncertainty.

Hide