One of the main challenges in curbing the spread of coronoavirus disease 2019 (COVID-19) is rooted in the limited technical capacities of current diagnostic testing platforms. Qualitative reverse transcription polymerase chain reaction, the gold standard diagnostic method, requires dedicated machinery and can only simultaneously process a predefined number of samples. Specimen pooling has been introduced to improve throughput, with recovery methods either requiring reiterative testing or failing to score the viral load. The proposed technology implements learned factor graph algorithms to upgrade the traditional group-testing (GT) recovery methods, to enable accurate, reliable and quantitative one-shot recovery of COVID-19-positive samples in a pooled specimen lot.
Attempts to contain the ongoing coronavirus disease 2019 (COVID-19) pandemic have been focusing on early identification and quarantine of infected individuals. The main diagnostic tool for COVID-19 carriage and estimate of viral load is based on viral RNA extraction and its qualitative detection using standard real-time reverse transcription polymerase chain reactions (rRT-PCR). The central bottlenecks associated with this technique are the limited number of samples that can be simultaneously processed in a single machine and the length of the process. To shorten the turnaround time, specimen pooling has been introduced, where several samples from different patients are mixed and processed as one. Several schemes were proposed for the identification of the positive samples within the pool. These recovery procedures are primarily based on either testing (GT) theory, which addresses group detection problems, or on compressed sensing (CS), which focuses on the recovery of sparse signals from compressed data. GT provides binary variables, indicating whether the individual is infected or not, and generally requires fewer sequential testing than CS. On the other hand, CS provides a real value, indicative of viral load. Therefore, there is a need to combine current methods so that a one-shot pooling operation will indicate for each sample if it is positive/negative, along with its viral load, and will also account for effects of noise and distortion induced in the RT-qPCR procedure.
This technology enables quantitative pooled sample recovery in a one-shot operation by extending the classical GT algorithm and defining a dedicated testing matrix, to enable elimination of samples that are definitely negative and quantitative analysis of the residual information.
The proposed strategy builds upon learned factor graphs, a family of inference algorithms which exploits a-priori knowledge of a relationship between underlying factors, and deep learning, to compute function nodes without having to explicitly model them. Thereafter, it carries out a sum-product operation to infer discrete values and assign them to each infected specimen. The approach has been used for symbol recovery in communications arenas and is applied here to enable identification of the subset of infected items and the viral load associated with each sample.
- Limited computational burden
- Suitable for various pooling combinations
- Facilitates identification of preferred pooling mappings
- A single learned function node can be trained using all available training sets
- One-shot inference
- Improved recovery guarantee
The ongoing COVID-19 pandemic has already taken a heavy toll on the global economy, public healthcare and education systems and personal physical and mental health, and threatens to have far-reaching impacts. Ramping up screening capacities has been argued a means of enabling safe return to normal activities, international travel and economic recovery. Large-scale testing to prevent or curb further outbreaks is of essence in all countries currently battling its spread, particularly in those on the verge of collapsing healthcare infrastructures.