In this issue of Blood, Nagata et al show for the first time that machine learning (ML) algorithms can discern patterns and identify diagnostically and prognostically relevant associations between genetic variants and cytomorphological changes in myelodysplastic syndromes (MDS), correlations that are too subtle or complex for the human eye to observe.1 

Nagata et al collected morphologic, clinical, and genetic data of 1079 MDS patients. Consensus clustering was used to group the patients based on their morphologic profiles. High-risk patients mainly segregated into 1 cluster; low-risk patients were represented by 4 clusters. Analysis of the genetic profiles by ML techniques revealed different genetic signatures with prognostic value for high-risk patients and identified significant correlations between genetic variants and clinical features for the morphological profiles of low-risk patients. IPSS-R, Revised International Prognostic Scoring System; P, patient.

Nagata et al collected morphologic, clinical, and genetic data of 1079 MDS patients. Consensus clustering was used to group the patients based on their morphologic profiles. High-risk patients mainly segregated into 1 cluster; low-risk patients were represented by 4 clusters. Analysis of the genetic profiles by ML techniques revealed different genetic signatures with prognostic value for high-risk patients and identified significant correlations between genetic variants and clinical features for the morphological profiles of low-risk patients. IPSS-R, Revised International Prognostic Scoring System; P, patient.

We know about many phenotype-genotype correlations in hematology, including PML-RARA and the classic hypergranular abnormal promyelocytes of acute promyelocytic leukemia, germline GATA2 mutations, and myelokathexis. In MDS, examples include the megakaryocyte dysmorphology in isolated del(5q) or MDS with ring sideroblasts and SF3B1 mutations.2  But what about for MDS in general?

Artificial intelligence (AI) is the umbrella term to describe any “nonhuman” that mimics cognitive skills like problem solving and learning. Together with “deep learning” and ML, AI is currently a buzzword across medical specialties and scientific disciplines, with its potential and risks hotly debated. We lack a clear picture yet of what AI and ML will mean for the future doctor-patient relationship.

We should remember that for at least 175 years, since Rudolf Virchow invented the name leuk-emia,3  hematological diagnostics have been based almost entirely on cell counting and cell phenotype, assessed using the light microscope. Until recently, all studies of prognostic markers and therapies in hematological malignancies have been undertaken using disease entities defined by microscopy. Yet, diagnostic ambiguity remains a regular occurrence. I have been reminded again and again by sometimes painful personal experience, many years of cooperation with colleagues in the context of proficiency testing, and joint microscopy sessions with John M. Bennett and many other key opinion leaders in morphology that morphological diagnosis of MDS in particular is often very difficult, with intra- and interobserver reproducibility lower than we would like.

Still, when experienced and skilled hematopathologists combine multiple factors such as blast counts, dysplasia assessment, enumeration of monocytes, ring sideroblasts, assessment of cellularity, and other features such as conventional cytogenetics, a sufficiently accurate diagnosis can be made most of the time. As blood and marrow sample numbers increase due to aging populations and morphologists are placed under increasing strain, we are urgently trying to integrate different methods, such as molecular genetics and immunophenotyping, to improve diagnosis and establish correlations that might eventually lead to improved targeted therapy options for patients.2,4 

Only in the last few decades have methods of genetics entered routine diagnostics in such a way that for the first time we have been able to investigate correlations between the old “art” of morphology and more recently discovered cytogenetic and molecular genetic changes defining disease states. Publications on molecular genetics in MDS date back <10 years.5,6 

Nagata et al have now linked well-established phenotypic methods (cytomorphology) and clinical features with the current molecular genetic approaches, including whole-exome sequencing and targeted sequencing, and have subject them to ML (see figure). The obtained results are therefore not independent of the gold-standard analyses, but offer the possibility of greater reliability. By unsupervised consensus clustering, the authors identified 5 distinct MDS morphological profiles with unique clinical characteristics, separating patients with different prognoses. The lower-risk patients in the Nagata et al series (reflected by 4 of these morphological profiles) were classified in 8 genetic signatures, associated with specific morphological profiles by Bayesian ML techniques. However, for validation, it is crucial not only that correlations between orthogonal data from phenotype and genotype are depicted, but also that new correlations are found and subgroups with clinical impact are identified as being associated with certain prognosis.

The goal of such a combined approach is to go beyond the current World Health Organization (WHO) MDS classification. So far, only SF3B1 was included in the WHO classification of 2017, so there is plenty of space for improvement on the molecular level.2  The approach used here will likely lead to new diagnostic and prognostic subcategories.

In MDS, dozens of genes are recurrently mutated and allelic heterogeneity is considerable, and with the exception of SF3B1 and ring sideroblasts, single-gene mutations have not been tightly linked to a specific phenotype.2  This makes it even more important to use tools such as ML to make the multidimensionality of molecular data sets and the great complexity of phenotypic characteristics conceivable and usable in the future.

In fact, an approach based on the gold standard of morphology and classical molecular genetics (eg, panel testing) is currently only an interim stage. The eyes of the experienced morphologist might miss features that AI might recognize. Further, individual genes do not paint the picture of a genome, but the latter may be used in the near future. Costs per genome (30×) are already <$1000. Meaningful combination of phenotypic and genotypic data will have the power to completely rewrite diagnostics in the next 10 years, and not only in hematology.

When an experienced morphologist reads a bone marrow smear, he or she “sees beyond” what an inexperienced morphologist observes. How does that work in the morphologist’s brain? A lot of practice, experience, and talent allow development of links between the art of observation with natural human intelligence. AI can accumulate practice and experience much faster than humans, so it is not surprising that it can go beyond what humans can accomplish.

AI-driven tools have found their way into our private environments. We live with talking devices such as Alexa or Siri, driver assistance technology, and language translation programs. Certainly these tools will soon be a regular part of the diagnostics and therapy of our patients. The medical knowledge around us is exploding, while the comprehension ability of human brains remains the same as it has been for centuries.

If we want to treat our patients correctly in the future, we must use all the tools available to us. We must test, validate, improve, and approve them.7  ML is part of tomorrow’s medicine, so we have to start today to challenge and develop it and to develop ourselves.8,9  To paraphrase Antonio Di leva,10  “AI will not replace physicians. However, physicians who use AI will replace those who don’t.”

Conflict-of-interest disclosure: T.H. declares no competing financial interests.

REFERENCES

REFERENCES
1.
Nagata
Y
,
Zhao
R
,
Awada
H
, et al
.
Machine learning demonstrates that somatic mutations imprint invariant morphologic features in myelodysplastic syndromes
.
Blood
.
2020
;
136
(
20
):
2249
-
2262
.
2.
Swerdlow
SH
,
Campo
E
,
Harris
NL
, et al
.
WHO Classification of Tumours of Haematopoietic and Lymphoid Tissues
. 4th ed.
Lyon: International Agency for Research on Cancer (IARC)
;
2017
.
3.
Virchow
R
.
Weisses Blut
.
Frorieps Notizen
.
1845
;
36
:
151
-
156
.
4.
Malcovati
L
,
Hellström-Lindberg
E
,
Bowen
D
, et al;
European Leukemia Net
.
Diagnosis and treatment of primary myelodysplastic syndromes in adults: recommendations from the European LeukemiaNet
.
Blood
.
2013
;
122
(
17
):
2943
-
2964
.
5.
Papaemmanuil
E
,
Gerstung
M
,
Malcovati
L
, et al;
Chronic Myeloid Disorders Working Group of the International Cancer Genome Consortium
.
Clinical and biological implications of driver mutations in myelodysplastic syndromes
.
Blood
.
2013
;
122
(
22
):
3616
-
3627, quiz 3699
.
6.
Haferlach
T
,
Nagata
Y
,
Grossmann
V
, et al
.
Landscape of genetic lesions in 944 patients with myelodysplastic syndromes
.
Leukemia
.
2014
;
28
(
2
):
241
-
247
.
7.
Topol
EJ
.
High-performance medicine: the convergence of human and artificial intelligence
.
Nat Med
.
2019
;
25
(
1
):
44
-
56
.
8.
Rajkomar
A
,
Dean
J
,
Kohane
I
.
Machine learning in medicine
.
N Engl J Med
.
2019
;
380
(
14
):
1347
-
1358
.
9.
Rahwan
I
,
Cebrian
M
,
Obradovich
N
, et al
.
Machine behaviour
.
Nature
.
2019
;
568
(
7753
):
477
-
486
.
10.
Di Ieva
A
.
AI-augmented multidisciplinary teams: hype or hope?
Lancet
.
2019
;
394
(
10211
):
1801
.