Predicting Leukemia Relapses Through Time Series and Other Advanced Techniques.
Every three minutes, one person in the U.S. is diagnosed with leukemia or a similar blood cancer. And every nine minutes, someone in the U.S. dies from these diseases.
Indeed, leukemia and related conditions are devastatingly common, even in countries with advanced medical systems. The diseases comprised around 10 percent of new cancer cases in the U.S. in 2021. More than 1.5 million people across the country either live with or are in remission from leukemia, lymphoma, myeloma, myelodysplastic syndromes (MDS), or myeloproliferative neoplasms (MPNs).
Despite these numbers, those afflicted with leukemia or other blood cancers have a better chance of survival now than ever before. The five-year survival rate for leukemia has more than quadrupled from 1960 to 2016.
Leukemia relapse: An all-too-common occurrence
Relapses among leukemia victims, however, are exceedingly common. Adults with acute lymphoblastic leukemia (ALL), for example, have a 50 percent chance of relapsing. Between 10 and 20 percent of ALL sufferers of all ages will relapse.
Relapses of other types of leukemia, such as acute myeloid leukemia (AML), also happen relatively frequently. That’s partially because chemotherapy doesn’t kill the leukemia stem cells that cause AML. “Leukemia stem cells tend to survive chemotherapy, stay in the bone marrow and regrow the disease,” says Dr. Jean Wang, Clinician Scientist at the Princess Margaret Cancer Centre in Toronto.
A patient’s relapse risk is currently assessed through various data points, including molecular and cytogenetic tests or clinical/biological factors such as age and white blood cell count. But these techniques are notoriously slow and often don’t produce results in time to save the patient from an aggressive relapse.
Accurate, more timely assessments are necessary to avoid over- or under-treatment and, ultimately, to ensure positive health outcomes.
How time series and other new techniques have improved leukemia relapse prediction
Several novel approaches for predicting leukemia relapses have recently been developed, including DNA and genetic testing and the use of machine learning models fed by time series data.
Gene signature testing: The LSC17 Test
Dr. Wang of the Princess Margaret Cancer Centre and colleagues recently developed a new gene signature test for AML that delivers accurate results within one to two days.
Dr. Wang and her fellow researchers identified a signature of 17 genes – an “LSC17 score” – that can accurately predict a patient’s risk of relapse following chemotherapy. The test uses the NanoString platform to assess the patient and calculate an LSC17 score. The higher the score, the more likely that patient will relapse.
Because the test is administered at the time of diagnosis, it can inform the most appropriate treatment methods – including experimental therapy or stronger initial therapy – based on the patient’s likelihood of relapse. “Using it as a correlative test in clinical trials would be really valuable to identify drugs that can benefit high-risk patients,” says Dr. Wang.
DNA sequencing: The NGS-MRD biomarker
Researchers from the University of Utah, the University of Pennsylvania, and other institutions recently identified a biomarker that almost always predicts which ALL patients will relapse or need further treatment within three to 12 months of CAR-T cell therapy.
While CAR-T cell therapy leads to complete remission in around 80 percent of children and young adult patients, about half of these patients eventually relapse.
“This study demonstrates that the best biomarker described to date for determining risk of relapse at any given time throughout the first year after CAR-T cell therapy with tisagenlecleucel is NGS-MRD assessment of the marrow with a cutoff of >0 cells detected,” the researchers wrote in Blood Cancer Discovery in December 2021.
Researchers identified the biomarker after analyzing hundreds of clinical trial samples, with samples drawn for screening at 1-, 3-, 6-, 9-, and 12-month intervals.
Time-lapse predictions: Machine learning and other computational methods
Other researchers have taken a similar interval approach described above and paired it with various computational methods, including machine learning. In late 2021, Hoffmann Et al. designed a synthetic experiment that compared the accuracy of three computational methods – mechanistic models, generalized linear models, and deep neural networks – to predict acute myeloid leukemia relapse.
The analysis showed that long-short-term-memory (LSTM) neural networks perform well for predicting recurrence when frequent measurements are available. However, mechanistic models or statistical approaches using prior knowledge are more effective for situations with limited data.
“Our study indicates that the optimization of measurement schemes and clinical protocols is a promising strategy to improve the overall prediction accuracy without necessarily requiring more measurements per patient,” the researchers wrote. “In our predictions for AML recurrence, we could reach a level of accuracy of about 80% for the prognosis of relapse occurrence within two years after diagnosis.”
Time-lapse techniques are also used to predict relapse in childhood acute lymphoblastic leukemia. Yeoh Et al. used time-series gene expression profiling (GEPs) of bone marrow samples during remission-induction therapy to predict the likelihood of relapse.
The researchers measured time-series changes for eight days following diagnosis, a calculation known as Effective Response Metric (ERM-D8). During independent blinded tests, the metric identified a threefold increased risk of relapse in those with an unfavorable ERM-D8. The metric “significantly improved risk assignment by picking up an additional 23.9 percent of relapses in patients without high-risk features by genetics,” the authors wrote.