Comparison of three machine learning algorithms for classification of B-cell neoplasms using clinical flow cytometry data.

Multiparameter flow cytometry data is visually inspected by expert personnel as part of standard clinical disease diagnosis practice. This is a demanding and costly process, and recent research has demonstrated that it is possible to utilize artificial intelligence (AI) algorithms to assist in the interpretive process. Here we report our examination of three previously published machine learning methods for classification of flow cytometry data and apply these to a B-cell neoplasm dataset to obtain predicted disease subtypes. Each of the examined methods classifies samples according to specific disease categories using ungated flow cytometry data. We compare and contrast the three algorithms with respect to their architectures, and we report the multiclass classification accuracies and relative required computation times. Despite different architectures, two of the methods, flowCat and EnsembleCNN, had similarly good accuracies with relatively fast computational times. We note a speed advantage for EnsembleCNN, particularly in the case of addition of training data and retraining of the classifier.

Recommendations for using artificial intelligence in clinical flow cytometry.

Flow cytometry is a key clinical tool in the diagnosis of many hematologic malignancies and traditionally requires close inspection of digital data by hematopathologists with expert domain knowledge. Advances in artificial intelligence (AI) are transferable to flow cytometry and have the potential to improve efficiency and prioritization of cases, reduce errors, and highlight fundamental, previously unrecognized associations with underlying biological processes. As a multidisciplinary group of stakeholders, we review a range of critical considerations for appropriately applying AI to clinical flow cytometry, including use case identification, low and high risk use cases, validation, revalidation, computational considerations, and the present regulatory frameworks surrounding AI in clinical medicine. In particular, we provide practical guidance for the development, implementation, and suggestions for potential regulation of AI-based methods in the clinical flow cytometry laboratory. We expect these recommendations to be a helpful initial framework of reference, which will also require additional updates as the field matures.

Angioimmunoblastic T-cell lymphoma presenting with severe plasmacytosis mimicking plasma cell leukemia.

Peripheral blood smear (A) demonstrates increased numbers of plasma cells (representative cells indicated by arrows), (B) demonstrates polytypic nature of plasma cells.

Tyramide-conjugated DNA barcodes enable signal amplification for multiparametric CODEX imaging.

Multiparametric imaging allows researchers to measure the expression of many biomarkers simultaneously, allowing detailed characterization of cell microenvironments. One such technique, CODEX, allows fluorescence imaging of >30 proteins in a single tissue section. In the commercial CODEX system, primary antibodies are conjugated to DNA barcodes. This modification can result in antibody dysfunction, and development of a custom antibody panel can be very costly and time consuming as trial and error of modified antibodies proceeds. To address these challenges, we developed novel tyramide-conjugated DNA barcodes that can be used with primary antibodies via peroxidase-conjugated secondary antibodies. This approach results in signal amplification and imaging without the need to conjugate primary antibodies. When combined with commercially available barcode-conjugated primary antibodies, we can very quickly develop working antibody panels. We also present methods to perform antibody staining using a commercially available automated tissue stainer and in situ hybridization imaging on a CODEX platform. Future work will include application of the combined tyramide-based and regular CODEX approach to image specific tumors with their immune cell infiltrates, including biomarkers that are currently difficult to image by regular CODEX.

Potential for Process Improvement of Clinical Flow Cytometry by Incorporating Real-Time Automated Screening of Data to Expedite Addition of Antibody Panels.

OBJECTIVES: We desired an automated approach to expedite ordering additional antibody panels in our clinical flow cytometry lab. This addition could improve turnaround times, decrease time spent revisiting cases, and improve consistency. METHODS: We trained a machine learning classifier to use our screening B-cell panel to predict whether we should order an additional panel to distinguish chronic lymphocytic lymphoma from mantle cell lymphoma. We used data from 2016 to 2018 for training and validation, and cases were restricted to the first case per patient (9,635 cases, 887 with the additional panel). We applied the model in real time over approximately 2.5 months in 2020 to 376 sequential cases, with automated email notifications for positive predictions. RESULTS: Using 80% of the data from 2016 to 2018 to train and 20% for validation, we achieved 95% area under the receiving operating characteristic curve (AUROC) and 94% accuracy in the validation set. Applying the classifier in real time achieved 89% AUROC and 94% real-time prediction accuracy (precision [positive predictive value] = 51%, recall [sensitivity] = 78%, and F1 score = 0.62). Fourteen of the 17 false positives had prior diagnoses to which the algorithm was not privy. CONCLUSIONS: As an observational, not interventional study, our system performed well on testing within our laboratory for identifying cases to be flagged but cannot be used without laboratory-specific modifications.

De Novo Identification and Visualization of Important Cell Populations for Classic Hodgkin Lymphoma Using Flow Cytometry and Machine Learning.

OBJECTIVES: Automated classification of flow cytometry data has the potential to reduce errors and accelerate flow cytometry interpretation. We desired a machine learning approach that is accurate, is intuitively easy to understand, and highlights the cells that are most important in the algorithm's prediction for a given case. METHODS: We developed an ensemble of convolutional neural networks for classification and visualization of impactful cell populations in detecting classic Hodgkin lymphoma using two-dimensional (2D) histograms. Data from 977 and 245 clinical flow cytometry cases were used for training and testing, respectively. Seventy-eight nongated 2D histograms were created per flow cytometry file. Shapley additive explanation (SHAP) values were calculated to determine the most impactful 2D histograms and regions within histograms. SHAP values from all 78 histograms were then projected back to the original cell data for gating and visualization using standard flow cytometry software. RESULTS: The algorithm achieved 67.7% recall (sensitivity), 82.4% precision, and 0.92 area under the receiver operating characteristic. Visualization of the important cell populations for individual predictions demonstrated correlations with known biology. CONCLUSIONS: The method presented enables model explainability while highlighting important cell populations in individual flow cytometry specimens, with potential applications in both diagnosis and discovery of previously overlooked key cell populations.

Hospitalization and mortality associated with SARS-CoV-2 viral clades in COVID-19.

The COVID-19 epidemic of 2019-20 is due to the novel coronavirus SARS-CoV-2. Following first case description in December, 2019 this virus has infected over 10 million individuals and resulted in at least 500,000 deaths world-wide. The virus is undergoing rapid mutation, with two major clades of sequence variants emerging. This study sought to determine whether SARS-CoV-2 sequence variants are associated with differing outcomes among COVID-19 patients in a single medical system. Whole genome SARS-CoV-2 RNA sequence was obtained from isolates collected from patients registered in the University of Washington Medicine health system between March 1 and April 15, 2020. Demographic and baseline clinical characteristics of patients and their outcome data including their hospitalization and death were collected. Statistical and machine learning models were applied to determine if viral genetic variants were associated with specific outcomes of hospitalization or death. Full length SARS-CoV-2 sequence was obtained 190 subjects with clinical outcome data. 35 (18.4%) were hospitalized and 14 (7.4%) died from complications of infection. A total of 289 single nucleotide variants were identified. Clustering methods demonstrated two major viral clades, which could be readily distinguished by 12 polymorphisms in 5 genes. A trend toward higher rates of hospitalization of patients with Clade 2 infections was observed (p = 0.06, Fisher's exact). Machine learning models utilizing patient demographics and co-morbidities achieved area-under-the-curve (AUC) values of 0.93 for predicting hospitalization. Addition of viral clade or sequence information did not significantly improve models for outcome prediction. In summary, SARS-CoV-2 shows substantial sequence diversity in a community-based sample. Two dominant clades of virus are in circulation. Among patients sufficiently ill to warrant testing for virus, no significant difference in outcomes of hospitalization or death could be discerned between clades in this sample. Major risk factors for hospitalization and death for either major clade of virus include patient age and comorbid conditions.

Creating Virtual Hematoxylin and Eosin Images using Samples Imaged on a Commercial CODEX Platform.

Multiparametric fluorescence imaging through CODEX allows the simultaneous imaging of many biomarkers in a single tissue section. While the digital fluorescence data thus obtained can provide highly specific characterizations of individual cells and microenvironments, the images obtained are different from those usually interpreted by pathologists (i.e., hematoxylin and eosin [H&E] slides and 3,3'-diaminobenzidine-stained immunohistochemistry slides). Having the fluorescence data plus coregistered H&E or similar data could facilitate the adoption of multiparametric imaging into regular workflows, as well as facilitate the transfer of algorithms and machine learning previously developed around H&E slides. Since commercial CODEX instruments do not produce H&E-like images by themselves, we developed a staining protocol and associated image processing to make "virtual H&E" images that can be incorporated into the CODEX workflow. While there are many ways to achieve virtual H&E images, including the use of a fluorescent nuclear stain and tissue autofluorescence to simulate eosin staining, we opted to combine fluorescent nuclear staining (through 4',6-diamidino-2-phenylindole) with actual eosin staining. We also output images derived from fluorescent nuclear staining and autofluorescence images for additional evaluation.

Profiling of immune dysfunction in COVID-19 patients allows early prediction of disease progression.

With a rising incidence of COVID-19-associated morbidity and mortality worldwide, it is critical to elucidate the innate and adaptive immune responses that drive disease severity. We performed longitudinal immune profiling of peripheral blood mononuclear cells from 45 patients and healthy donors. We observed a dynamic immune landscape of innate and adaptive immune cells in disease progression and absolute changes of lymphocyte and myeloid cells in severe versus mild cases or healthy controls. Intubation and death were coupled with selected natural killer cell KIR receptor usage and IgM+ B cells and associated with profound CD4 and CD8 T-cell exhaustion. Pseudo-temporal reconstruction of the hierarchy of disease progression revealed dynamic time changes in the global population recapitulating individual patients and the development of an eight-marker classifier of disease severity. Estimating the effect of clinical progression on the immune response and early assessment of disease progression risks may allow implementation of tailored therapies.

Longitudinal immune profiling of mild and severe COVID-19 reveals innate and adaptive immune dysfunction and provides an early prediction tool for clinical progression.

With a rising incidence of COVID-19-associated morbidity and mortality worldwide, it is critical to elucidate the innate and adaptive immune responses that drive disease severity. We performed longitudinal immune profiling of peripheral blood mononuclear cells from 45 patients and healthy donors. We observed a dynamic immune landscape of innate and adaptive immune cells in disease progression and absolute changes of lymphocyte and myeloid cells in severe versus mild cases or healthy controls. Intubation and death were coupled with selected natural killer cell KIR receptor usage and IgM+ B cells and associated with profound CD4 and CD8 T cell exhaustion. Pseudo-temporal reconstruction of the hierarchy of disease progression revealed dynamic time changes in the global population recapitulating individual patients and the development of an eight-marker classifier of disease severity. Estimating the effect of clinical progression on the immune response and early assessment of disease progression risks may allow implementation of tailored therapies.

Outcomes associated with SARS-CoV-2 viral clades in COVID-19.

Background The COVID-19 epidemic of 2019-20 is due to the novel coronavirus SARS-CoV-2. Following first case description in December, 2019 this virus has infected over 10 million individuals and resulted in at least 500,000 deaths world-wide. The virus is undergoing rapid mutation, with two major clades of sequence variants emerging. This study sought to determine whether SARS-CoV-2 sequence variants are associated with differing outcomes among COVID-19 patients in a single medical system. Methods Whole genome SARS-CoV-2 RNA sequence was obtained from isolates collected from patients registered in the University of Washington Medicine health system between March 1 and April 15, 2020. Demographic and baseline medical data along with outcomes of hospitalization and death were collected. Statistical and machine learning models were applied to determine if viral genetic variants were associated with specific outcomes of hospitalization or death. Findings Full length SARS-CoV-2 sequence was obtained 190 subjects with clinical outcome data. 35 (18.4%) were hospitalized and 14 (7.4%) died from complications of infection. A total of 289 single nucleotide variants were identified. Clustering methods demonstrated two major viral clades, which could be readily distinguished by 12 polymorphisms in 5 genes. A trend toward higher rates of hospitalization of patients with Clade 2 was observed (p=0.06). Machine learning models utilizing patient demographics and co-morbidities achieved area-under-the-curve (AUC) values of 0.93 for predicting hospitalization. Addition of viral clade or sequence information did not significantly improve models for outcome prediction. Conclusion SARS-CoV-2 shows substantial sequence diversity in a community-based sample. Two dominant clades of virus are in circulation. Among patients sufficiently ill to warrant testing for virus, no significant difference in outcomes of hospitalization or death could be discerned between clades in this sample. Major risk factors for hospitalization and death for either major clade of virus include patient age and comorbid conditions.

Characterization of bilirubin interference in three commonly used digoxin assays.

BACKGROUND: Due to the narrow therapeutic range of digoxin, determining serum/plasma digoxin concentrations is critical for assessing patients with congestive heart failure, atrial fibrillation, and certain types of arrhythmias. However, digoxin quantification by competitive immunoassays is susceptible to interferences that may alter the accuracy of its measurement in patient plasma. This study aimed to characterize the extent of bilirubin interference in three commonly used digoxin immunoassays. METHODS: Digoxin concentrations were compared using the Beckman Coulter® Unicel DxI 800, the Vitros® 4600, and the Roche Cobas® 8000 in neat or digoxin-spiked icteric and non-icetric plasma samples. A mixing study was performed to demonstrate how digoxin quantification is affected by bilirubin. An equation was derived that predicts the response of the DxI 800, given known bilirubin and digoxin concentrations. RESULTS: The DxI reported detectable concentrations of digoxin in high bilirubin samples with no added digoxin, while the Vitros® 4600 and Cobas® 8000 gave virtually undetectable results. Spiking digoxin into samples with elevated bilirubin concentrations resulted in a higher percent recovery for the DxI 800 when compared to the other two platforms. The mixing study also revealed an increase in the percent recovery in the DxI 800, while the Vitros® 4600 and Cobas® 8000 were comparable to the expected concentration of digoxin. CONCLUSIONS: The DxI 800 is most prone to interference by bilirubin, while the Vitros® 4600 and Cobas® 8000 are relatively unaffected. Icteric samples should be interpreted with caution if digoxin quantification is needed, especially on the DxI 800 assay.

Macrophage activation syndrome in a patient with axial spondyloarthritis on adalimumab.

Macrophage activation syndrome (MAS) is a rare and potentially fatal condition characterized by excessive activation and uncontrolled proliferation of T lymphocytes and macrophages, leading to overwhelming systemic inflammation and cytokine release. MAS has been reported with viral infections, autoimmune disorders, malignancies, and medications. We describe a case of a patient with axial spondyloarthritis (axSpA) treated with adalimumab, who presented with MAS.

Acute Liver Failure Due to Echovirus 9 Associated With Persistent B-Cell Depletion From Rituximab.

We describe a case of fatal acute liver failure due to echovirus 9 in the setting of persistent B-cell depletion and hypogammaglobulinemia 3 years after rituximab therapy. Metagenomic next-generation sequencing further specified the etiologic agent. Early recognition may provide an opportunity for interventions including intravenous immunoglobulin and liver transplantation.

Using fixed fiduciary markers for stage drift correction.

To measure nanometric features with super-resolution requires that the stage, which holds the sample, be stable to nanometric precision. Herein we introduce a new method that uses conventional equipment, is low cost, and does not require intensive computation. Fiduciary markers of approximately 1 µm x 1 µm x 1 µm in x, y, and z dimensions are placed at regular intervals on the coverslip. These fiduciary markers are easy to put down, are completely stationary with respect to the coverslip, are bio-compatible, and do not interfere with fluorescence or intensity measurements. As the coverslip undergoes drift (or is purposely moved), the x-y center of the fiduciary markers can be readily tracked to 1 nanometer using a Gaussian fit. By focusing the light slightly out-of-focus, the z-axis can also be tracked to < 5 nm for dry samples and <17 nm for wet samples by looking at the diffraction rings. The process of tracking the fiduciary markers does not interfere with visible fluorescence because an infrared light emitting diode (IR-LED) (690 and 850 nm) is used, and the IR-light is separately detected using an inexpensive camera. The resulting motion of the coverslip can then be corrected for, either after-the-fact, or by using active stabilizers, to correct for the motion. We applied this method to watch kinesin walking with ≈ 8 nm steps.

Single-molecule-based super-resolution images in the presence of multiple fluorophores.

Several super-resolution techniques exist, yet most require multiple lasers, use either large or weakly emitting fluorophores, or involve chemical manipulation. Here we show a simple technique that exceeds the standard diffraction limit by 5-15× on fixed samples, yet allows the user to localize individual fluorophores from among groups of crowded fluorophores. It relies only on bright, organic fluorophores and a sensitive camera, both of which are commercially available. Super-resolution is achieved by subtracting sequential images to find the fluorophores that photobleach (temporarily or permanently), photoactivate, or bind to the structure of interest in transitioning from one frame to the next. These fluorophores can then be localized via Gaussian fitting with selective frame averaging to achieve accuracies much better than the diffraction limit. The signal-to-noise ratio decreases with the square root of the number of nearby fluorophores, producing average single-molecule localization errors that are typically <30 nm. Surprisingly, one can often extract signal when there are approximately 20 fluorophores surrounding the fluorophore of interest. Examples shown include microtubules (in vitro and in fixed cells) and chromosomal DNA.

Two-photon 3D FIONA of individual quantum dots in an aqueous environment.

We report the first two-photon (2P) microscopy of individual quantum dots (QDs) in an aqueous environment with both widefield and point-scan excitations at nanometer accuracy. Thiol-containing reductants suppress QD blinking and enable measurement of the 36 nm step size of individual Myosin V motors in vitro. We localize QDs with an accuracy of 2-3 nm in all three dimensions by using a 9 × 9 matrix excitation hologram and an array detector, which also increases the 3D scan imaging rate by 80-fold. With this 3D microscopy we validate the LamB receptor distribution on E. coli and the endocytosis of EGF-receptors in breast cancer cells.

Counting bungarotoxin binding sites of nicotinic acetylcholine receptors in mammalian cells with high signal/noise ratios.

Nicotinic acetylcholine receptors are some of the most studied synaptic proteins; however, many questions remain that can only be answered using single molecule approaches. Here we report our results from single α7 and neuromuscular junction type nicotinic acetylcholine receptors in mammalian cell membranes. By labeling the receptors with fluorophore-labeled bungarotoxin, we can image individual receptors and count the number of bungarotoxin-binding sites in receptors expressed in HEK 293 cells. Our results indicate that there are two bungarotoxin-binding sites in neuromuscular junction receptors, as expected, and five in α7 receptors, clarifying previous uncertainty. This demonstrates a valuable technique for counting subunits in membrane-bound proteins at the single molecule level, with nonspecialized optics and with higher signal/noise ratios than previous fluorescent protein-based techniques.