Applications and challenges of ML in the management of AML
| Application . | Cytomorphology/histology . | Immunophenotyping . | Clinical data . | Cytogenetics/molecular genetics . | New therapies/prognostic scores . |
|---|---|---|---|---|---|
| Improvements needed | Precision of image segmentation (eg, detection of cell boundaries) | MRD evaluation (eg, standardization of cutoffs to form a decision boundary) | Integration of different high-dimensional data sets | Availability of data | Prospective studies for validation |
| Number values such as laboratory results as well as written text are better evaluated by different ML techniques (integrative models needed) | Models trained on online data (eg, Beat-AML or The Cancer Genome Atlas) may not be accurate on regional data | Majority of ML studies are only retrospective; models have to be evaluated in a prospective manner to evaluate their translational application in patient care | |||
| Feature extraction (eg, relation of nucleus to cytoplasm) | Classification methods | Standardization of clinical reports | Pre-processing of high-dimensional data | ||
| Which combination of different ML techniques shows the most accurate results? | Uniformity of clinical reports (eg, with standardized vocabulary) makes natural language processing easier | Accurate filtering of biosignatures is needed before classification | |||
| Cell classification | |||||
| Labeling by field experts needed | |||||
| Outlook | • Integrated workflow of various ML techniques to guide clinical decision-making | ||||
| • Strict legal and regulatory framework to ensure patient safety | |||||
| • Prospective clinical trials to verify robustness of ML models | |||||
| • Physicians with basic knowledge in ML techniques to optimally implement ML into clinical practice | |||||
| Application . | Cytomorphology/histology . | Immunophenotyping . | Clinical data . | Cytogenetics/molecular genetics . | New therapies/prognostic scores . |
|---|---|---|---|---|---|
| Improvements needed | Precision of image segmentation (eg, detection of cell boundaries) | MRD evaluation (eg, standardization of cutoffs to form a decision boundary) | Integration of different high-dimensional data sets | Availability of data | Prospective studies for validation |
| Number values such as laboratory results as well as written text are better evaluated by different ML techniques (integrative models needed) | Models trained on online data (eg, Beat-AML or The Cancer Genome Atlas) may not be accurate on regional data | Majority of ML studies are only retrospective; models have to be evaluated in a prospective manner to evaluate their translational application in patient care | |||
| Feature extraction (eg, relation of nucleus to cytoplasm) | Classification methods | Standardization of clinical reports | Pre-processing of high-dimensional data | ||
| Which combination of different ML techniques shows the most accurate results? | Uniformity of clinical reports (eg, with standardized vocabulary) makes natural language processing easier | Accurate filtering of biosignatures is needed before classification | |||
| Cell classification | |||||
| Labeling by field experts needed | |||||
| Outlook | • Integrated workflow of various ML techniques to guide clinical decision-making | ||||
| • Strict legal and regulatory framework to ensure patient safety | |||||
| • Prospective clinical trials to verify robustness of ML models | |||||
| • Physicians with basic knowledge in ML techniques to optimally implement ML into clinical practice | |||||