A machine learning and centrifugal microfluidics platform for bedside prediction of sepsis

Abstract

Sepsis is a life-threatening organ dysfunction due to a dysfunctional response to infection. Delays in diagnosis have substantial impact on survival. Herein, blood samples from 586 in-house patients with suspected sepsis are used in conjunction with machine learning and cross-validation to define a six-gene expression signature of immune cell reprogramming, termed Sepset, to predict clinical deterioration within the first 24 h (h) of clinical presentation. Prediction accuracy (~90% in early intensive care unit (ICU) and 70% in emergency room patients) is validated in 3178 patients from existing independent cohorts. A RT-PCR-based Sepset detection test shows a 94% sensitivity in 248 patients to predict worsening of the sequential organ failure assessment scores within the first 24 h. A stand-alone centrifugal microfluidic instrument that automates whole-blood Sepset classifier detection is tested, showing a sensitivity of 92%, and specificity of 89% in identifying the risk of clinical deterioration in patients with suspected sepsis.

Fig. 1. Overview of prediction classifier development, validation and analytical performance testing.

A Machine Learning, feature reduction and validation of the Sepset classifier. B Experimental Validation and Analytical Performance Testing of the Sepset Classifier on the PREDICT device.

Fig. 2. Microfluidic platform and cartridges.

A Microfluidic workflow for automation of RNA extraction and downstream ddPCR on the PREDICT platform and related cartridges. B PREDICT platform showing rotor (1), pneumatic manifold (2), cartridge installed on the rotor (3) and connected to the external PCR tube using world-to-chip tubing inserted into the wirelessly controlled platform heater (4). C RNA extraction cartridge design. D Image of the ddPCR cartridge design. E Micrograph showing close-up view of droplet generation chamber and nozzles. The scale bar is 2.4 mm. F Micrograph showing close-up view of droplet streams generated using the ddPCR cartridge. The scale bar is 2.4 mm. G Size distribution of droplet diameter. The inset shows optical micrograph of droplet monolayer using ddPCR cartridge. The scale bar is 150 µm. H Example of acquired fluorescence images showing droplet monolayer within a region of the imaging chamber. For clarity, only a zoomed-in portion of the imaging chamber region is shown to increase visibility of droplets. The scale bar is 400 µm. I Intensity maps for different fluorophores. Horizontal lines denote the threshold for positive and negative counts.

Fig. 3. Feature reduction and development of a gene classifier that predicts deterioration-risk-groups in patients.

This started with taking in-house RNA seq data from patient collected from a heterogenous cohort of patients with suspected sepsis (top left) to reduce our original published gene signature down to 6-genes (Sepset), for which expression could be related to 2 housekeeping genes. Feature selection was performed using machine learning (ML) and AI and the classifier validated in samples from published transcriptomic studies. Molecular assay is then developed by designing and testing primer/probe sequences specific to the target genes using digital droplet PCR. In parallel, sample-to-answer microfluidic platform and cartridges are developed (bottom right) and analytical performance of multiplex quantitative assay is tested. Prognostic enrichment is obtained by analyzing the results using ML algorithm to determine the percent likelihood of significant clinical deterioration within the immediate next 24 h. The deployment of PREDICT platform (center) at the point-of-care is anticipated to aid in triage and management of prospective sepsis within the first 3 h of clinical presentation.