Artificial intelligence in hematopoietic stem cell research and associated malignancies: From disease modeling to cell manufacturing

Table 1 Stem-cell relevance map linking each application area to stem-cell biological object, data modality, AI methodology, output, clinical/biological endpoint, and validation status

Application	Stem-cell object	Data modality	AI method/model	Dataset/scale	Key performance metrics	Clinical/biological endpoint	Validation status	Ref.
A: Normal HSC biology
HSC differentiation modeling	Normal HSCs, MPPs	scRNA-seq, scATAC-seq	VIA (Voyager) - lazy-teleporting MCMC trajectory inference	Human CD34+ hematopoiesis (multi-site)	F1 > 0.9 for rare lineage populations; robust across scRNA-seq and scATAC-seq	HSC differentiation dynamics; lineage bifurcation mapping	Cross-modal validation (scRNA-seq + scATAC-seq)	[5]
HSC aging - chromatin architecture	Young vs aged murine HSCs	3D confocal DAPI chromatin imaging	ChromAgeNet - CNN on 3D nuclear images	1229 HSC nuclei (551 young, 678 aged)	AUROC: 0.77 ± 0.03; accuracy 0.68 ± 0.05; outperforms handcrafted features (AUROC: 0.73)	Biological age prediction; epigenetic rejuvenation detection	5-fold cross-validation; drug-treatment validation	[6]
HSC division dynamics	HSCs/progenitors - 3 age groups (young, mid-life, aged)	Single-cell gene expression	ANN	9 stem/progenitor populations across 3 age groups	89% accuracy (cell type + donor age); 96% accuracy (regenerative status)	Age-dependent self-renewal; niche sensitivity across lifespan	Cross-validation	[10]
HSC morphological classification	HSCs vs MPPs	Bright-field microscopy images	Deep CNN (morphology-based)	Steady-state bright-field image dataset	High accuracy distinguishing HSCs from MPPs; morphological features encode functional state; > 98% accuracy in leukemic BM context	Non-destructive functional state identification without molecular profiling	Steady-state + leukemic BM validation	[11]
HSC quiescence regulation	LTHSCs, STHSCs	scRNA-seq integrated with niche signals (TPO, SCF, ANGPT1)	Boolean network modeling (constraint programming + model checking)	Pseudotrajectory-derived gene expression states	Predicted stable LTHSC, STHSC, and proliferating states; novel p53-ROS regulatory mechanism identified	Quiescence maintenance; stem cell pool regulation	Experimental validation of predicted mechanisms	[12]
Epigenetic/GRN regulation	HSCs at differentiation stages	scRNA-seq, scATAC-seq	SCENIC (TF network inference); CellOracle (in silico TF perturbation)	Multi-dataset hematopoietic single-cell data	Stage-specific TF activities (GATA1, CEBPD, IRF8) validated by chromatin accessibility	TF programs governing HSC self-renewal and differentiation	Cross-validated in mouse and human hematopoiesis	[13,14]
Multi-omic integration	HSCs/progenitors - incomplete or unpaired multi-omics	scRNA-seq + scATAC-seq + CITE-seq	scMaui (product-of-experts VAE); totalVI (joint RNA + protein latent space)	PBMC CITE-seq benchmarks; human HSPC datasets	Robust batch correction; missing-modality handling; joint RNA + protein embedding	Systems-level HSC state characterization; HSC/LSC surface phenotype + transcriptional integration	Benchmarked on standard CITE-seq datasets; applied to HSPC/LSC contexts	[15,16]
B: AML and LSCs
AML risk stratification	LSC-enriched compartments	Clinical, molecular, cytogenetic, and genomic features	Supervised ML (ensemble methods)	1383 patients - multicenter cohort	Predicted complete remission and 2-year OS; outperformed ELN-only stratification	Prognosis; LSC burden surrogate; treatment-intensity guidance	External validation in an independent cohort	[17]
AML transcriptomic risk/epigenetic subtypes	AML blast/LSC-enriched populations	Transcriptomic + epigenetic multi-omics (multi-center)	ML epigenetic subtype classification; prognostic integration	Multi-center transcriptomic cohort	Prognostic stratification by epigenetic subtype; integration with ELN 2022	Relapse risk; therapeutic vulnerability by epigenetic class	Multi-center validation	[18]
AML drug response prediction - multi-omics ensemble (MDREAM)	AML patient-derived blasts; LSC-enriched compartments	Multi-omics: Genomic mutations + transcriptomic gene expression (ex vivo drug sensitivity)	Ensemble ML (MDREAM framework; multi-omics integration)	BeatAML cohort: 278 training/183 validation; + Swedish AML cohort (n = 45); relapsed/refractory cohort (n = 12)	Spearman r = 0.68 (BeatAML validation); 77% correct responder identification at prediction confidence > 0.75	Drug response prediction in AML patients; therapy selection; resistance identification	External multi-cohort validation (BeatAML + 2 independent cohorts)	[19]
AML targeted therapy - network ML	LSC populations; genomic/proteomic networks	Genomic and proteomic datasets	Network-based machine learning	Multi-dataset genomic + proteomic	Identified resistance pathways and therapeutic vulnerabilities	Immunotherapy target prioritization; resistance mechanism mapping	Not fully specified	[20]
AML MRD detection (MAGIC-DR)	Residual LSC-like subpopulations (immature monocytic cells)	Multiparameter flow cytometry	XGBoost + UMAP (MAGIC-DR framework)	AML MRD specimen cohort	AUC of 0.97; strong concordance with expert manual analysis; identified LSC-like residual populations missed by manual gating	Relapse risk; MRD+/- classification; sub-threshold LSC-like disease detection	Prospective validation cohort	[21]
AML LSC phenotyping/MRD	CD34+CD38- LSC-enriched compartment	Flow cytometry (CD117, CD34, HLA-DR)	Random forest + SHAP explainability	194 patients (AML, AML-CR, normal BM)	Accuracy: 94.92%; AUC: 94.83%; CD117, CD34, HLA-DR top discriminators	LSC detection; early relapse identification; AML vs CR vs normal classification	Validated on a 194-patient clinical cohort	[22]
BCL-2/venetoclax response in LSC	LSC-enriched blasts	scRNA-seq + ex vivo drug sensitivity data	XGBoost	VenEx clinical trial patient samples	Venetoclax response prediction r = 0.71-0.84	LSC apoptotic dependency; treatment selection for LSC eradication	VenEx clinical trial validation	[23]
Menin inhibition - LSC clonal architecture	KMT2A-rearranged/NPM1-mutant LSCs	Single-cell multi-omics (scRNA-seq + scATAC-seq)	DL clonal architecture mapping	Clinical trial and patient-derived samples	Mapped HOXA/MEIS1 program activity; tracked LSC self-renewal disruption under Menin inhibition	LSC self-renewal disruption; Menin inhibitor response monitoring	Clinical trial context	[24,25]
LSC surface target prioritization	LSC populations (CD34+)	Multiparameter flow cytometry (LSC immunophenotyping)	ML ranking/feature prioritization	Multi-dataset LSC surface marker profiling	CD123, CLL-1 identified as top immunotherapy targets vs normal HSC	Immunotherapy target selection; CAR-T/bispecific antibody design	Informed clinical trial design	[26]
Synergistic drug combinations - LSC	Diagnosis/relapse LSC-enriched populations	Paired scRNA-seq + ex vivo drug screens	XGBoost (combination prediction)	Paired diagnosis/relapse patient samples	Synergistic drug pair identification; clinically actionable 2-week turnaround	Personalized relapse therapy; LSC-selective combination design	Flow cytometry validation	[23]
AML explainable AI/precision oncology	AML patients (bulk + LSC-enriched contexts)	Transcriptomic, clinical, molecular	Explainable ML: SHAP, attention, ensemble XAI	Multi-center AML cohorts	Improved clinical interpretability; synergistic drug-response signatures via ensemble XAI	Clinical decision support; treatment selection; regulatory compliance	Multi-center validation	[27-29]
C: MMSCs
MM spatial mapping - bone marrow niche	MMSCs; niche cells (BLIMP1+, CD8+, stromal)	Multiplex-IHC trephine biopsies	DL: MoSaicNet (tissue segmentation) + AwareNet (rare-cell detection)	MGUS and NDMM clinical trephine biopsy cohort	Spatial heterogeneity as key MGUS-to-NDMM distinction; tumor-immune spatial proximity mapped	Sanctuary identification for quiescent MMSCs; immune exclusion mapping; MGUS vs NDMM distinction	Validated on clinical MGUS and NDMM samples	[30]
MMSC niche interactions - treatment resistance	Quiescent MMSCs; BM stromal cells	Multi-omics datasets	Neural networks/DL	Multi-omics patient-derived datasets	Prediction of treatment resistance linked to niche-protective microenvironments	Niche-mediated drug resistance; MMSC persistence	Experimental/in vitro validation (details not fully specified)	[31,32]
D: Drug discovery and LSC-targeted therapy
Virtual screening/AI-enhanced docking	LSC-targeted compounds; kinase/epigenetic targets	Multi-billion-compound chemical libraries (structure-based)	AI-enhanced virtual screening; molecular docking; structure-based drug design	Ultra-large compound libraries	Accelerated hit/Lead identification; 11 confirmed hits validated by X-ray crystallography; reduced screening timelines	LSC-targeted drug identification; multi-target resistance-overcoming agents	Experimental validation (X-ray crystallography of confirmed hits)	[7]
Drug repurposing - AML	AML/LSC targets (polypharmacological)	FDA-approved compound structures; AML molecular targets	Structure-based virtual screening; molecular dynamics simulation	FDA-approved drug library	Identified repurposable compounds with polypharmacological AML/LSC activity	Rapid clinical translation; overcoming LSC therapy resistance via repurposing	In silico validation with molecular dynamics	[33]
E: Biomanufacturing and cell therapy production
HSC product quality control	CD34+ HSC products	Bright-field microscopy (manufacturing line)	Deep CNN (computer vision)	Clinical HSC manufacturing image dataset	High accuracy for cell viability and phenotype classification; validated vs flow cytometry	Real-time manufacturing QC; batch release decisions	Validated against the flow cytometry gold standard	[34]
CD34 yield prediction - cord blood	Cord blood HSCs (CD34+)	Pre/post-processing cell counts; CBU characteristics	Back-propagation ANN	802 CBUs	56.99% prediction accuracy for CD34 dose	Graft potency estimation; transplant planning	Validated on 802 CBUs	[35]
PBSC apheresis yield prediction	Autologous/allogeneic PBSC donors	Pre-apheresis CD34 counts; donor clinical characteristics	ML regression	Multi-site clinical apheresis dataset	Accurate harvest outcome prediction supporting collection scheduling	Collection efficiency; donor scheduling optimization	Clinical validation	[36]
Ex vivo HSC expansion optimization	HSCs in bioreactor culture	Multi-parameter sensors (O2, glucose, lactate, cytokine)	Reinforcement learning + digital twin modeling	Proof-of-concept bioreactor dataset	Optimized culture conditions; improved expansion yield via adaptive control	Expansion efficiency; GMP process optimization	Proof-of-concept validation	[37]
CAR-T manufacturing - smart hospital	CAR-T cell products (autologous)	Process sensor data; imaging; manufacturing records	AI + automation (Industry 4.0)	Pilot smart manufacturing hospital data	Reduced vein-to-vein time; standardized product quality	Scalable autologous CAR-T production; accessibility	Pilot implementation validation	[38]
Digital twin/Bioprocessing 4.0	Cell therapy products; HSC/CAR-T	Multi-parameter bioprocess sensor streams	Digital twin simulation; reinforcement learning; AI-assisted process control	Pharma manufacturing simulation datasets	Real-time process optimization; predictive quality control	GMP compliance; scalable biomanufacturing	Simulation and proof-of-concept	[8,9,39]
F: HSCT - transplantation, HLA matching, and GVHD
HLA genotyping extraction (NLP)	HSCT donors and recipients	Electronic health records (free text; 70000+ HLA reports)	Rule-based NLP (Python regex; 124 extraction + 8 cleaning rules)	70000+ HLA reports (Seoul National University Hospital)	Precision: 0.892-0.999; recall: 0795-0.998 across HLA-A, B, C, DR, DQ	Donor-recipient HLA matching; adverse drug reaction prediction	Clinical validation at the SNUH registry	[40]
Acute GVHD prediction (CNN-NLP)	HSCT recipients	Clinical notes + HLA typing data	CNN-NLP hybrid (word2vec encoding of HLA antigens/alleles)	18763 patients (Japanese Transplant Registry)	Stratified cumulative incidence 318% (low-risk) to 54.8% (high-risk); superior to Cox models	aGVHD risk stratification; immunosuppression planning	Registry-based validation (18763 patients)	[41]
Chronic GVHD phenotyping (ML + NLP)	HSCT patients with cGVHD	Clinical notes (organ involvement: Mouth, eye, liver, GI, joints, fascia, skin)	ML feature extraction + NLP clinical narrative analysis	Multicenter HSCT patient cohort	7 distinct cGVHD phenotypes; 2.24-fold mortality difference (high vs low risk); independent of NIH severity criteria	Mortality stratification; personalized immunosuppression	Multi-center clinical validation	[42]
Comprehensive HSCT AI care	HSCT recipients (full care pathway)	Transplant documentation; clinical narratives; registry data	NLP + ML (composite AI review)	Multi-institutional registry and EHR data	Automated complication parsing; infection management; NRM prediction	Early warning systems; donor selection; complication management	Multi-institutional review	[4]

AI: Artificial intelligence; HSC: Hematopoietic stem cell; MPP: Multipotent progenitor; scRNA-seq: Single-cell RNA sequencing; scATAC-seq: Single-cell assay of transposase-accessible chromatin sequencing; VIA: Voyager in single-cell RNA sequencing analysis; MCMC: Markov chain Monte Carlo; ADAPI: 4’,6-diamidino-2-phenylindole; CNN: Convolutional neural network; AUROC: Area under the receiver operating characteristic curve; ANN: Artificial neural network; BM: Bone marrow; LTHSC: Long-term hematopoietic stem cell; STHSC: Short-term hematopoietic stem cell; TPO: Thrombopoietin; SCF: Stem cell factor; ANGPT1: Angiopoietin-1; ROS: Reactive oxygen species; GRN: Gene regulatory network; SCENIC: Single-cell regulatory network inference and clustering; TF: Transcription factor; CITE-seq: Cellular indexing of transcriptomes and epitopes by sequencing; VAE: Variational autoencoder; PBMC: Peripheral blood stem cell; LSC: Leukemic stem cell; HSPC: Hematopoietic stem and progenitor cell; AML: Acute myeloid leukemia; ML: Machine learning; OS: Overall survival; ELN: European Leukemia Net; MRD: Measurable residual disease; XGBoost: Extreme Gradient Boosting; UMAP: Uniform manifold approximation and projection; AUC: Area under the curve; SHAP: SHapley Additive exPlanations; AML-CR: Acute myeloid leukemia complete remission; CR: Complete remission; DL: Deep learning; CLL-1: C-type lectin-like molecule-1; CAR-T: Chimeric antigen receptor T-cell; XAI: Explainable artificial intelligence; MMSC: Multiple myeloma stem cell; MM: Multiple myeloma; MGUS: Monoclonal gammopathy of undetermined significance; NDMM: Newly diagnosed multiple myeloma; FDA: Food and Drug Administration; QC: Quality control; CBU: Cord blood unit; GMP: Good Manufacturing Practice; HSCT: Hematopoietic stem cell transplantation; GVHD: Graft-vs-host disease; NLP: Natural language processing; HLA: Human leukocyte antigen; aGVHD: Acute graft-vs-host disease; cGVHD: Chronic graft-vs-host disease; NRM: Non-relapse mortality.

Table 2 Public datasets and computational tools specifically applicable to hematopoietic stem-cell artificial intelligence research

Stem-cell anchor	Resource name	Resource type	What it’s for (HSC/LSC-specific)	Access/Ref.
A: Reference atlases for HSC → lineage differentiation (where “stemness” lives)
Human bone-marrow baseline map for projecting new normal/AML datasets and interpreting “where the cell sits” on the HSC → lineage hierarchy	BoneMarrowMap	scRNA-seq atlas + projection tool	Reference for CD34+ HSPC balance + mature compartments; projection/classification of new hematopoietic or leukemic cells	GitHub/[90]
Canonical mouse HSPC differentiation landscape used as benchmark for TI and lineage priming	Nestorowa et al[91] Mouse HSPC atlas	scRNA-seq	Mouse HSPC heterogeneity + lineage trajectories; common benchmark dataset	GEO GSE81682/[91]
Early-life/developmental hematopoiesis context	Perinatal bone marrow development atlas	scRNA-seq	Niche development + marrow colonization programs	[92,93]
Cord blood HSC heterogeneity (neonatal stemness programs; ex vivo culture/transplant relevance)	Human cord blood HSC populations	scRNA-seq	CD34pos vs CD34neg HSC populations; neonatal HSC state differences	GEO GSE237832/[94]
Human HSC activation states (quiescence → activation continuum)	Human HSC activation trajectory	scRNA-seq (SMART-seq2)	Activation trajectories in the primitive CD34+CD38-CD45RA- compartment	EGA study1
Integrated RNA + chromatin programs across hematopoiesis (epigenetic stemness + TF logic)	Murine HSPC multimodal atlas	scRNA-seq + scATAC-seq	Joint transcriptomic/epigenetic differentiation programs	[95]
B: Multimodal phenotype ↔ transcriptome (HSC/LSC immunophenotype as a first-class signal)
Method benchmark for RNA + protein integration (not HSC-specific, but enables HSC surface-marker integration)	PBMC CITE-seq (10X genomics)	CITE-seq (RNA + protein)	Benchmarking multimodal integration models, you then apply them to marrow HSPC/LSC	10X datasets referenced as common benchmarks in the totalVI context[15]
High-parameter immune niche cell reference (HSC niche signals/immune interactions)	Murine spleen/lymph node CITE-seq	CITE-seq	Large protein panel; good demonstration dataset for multimodal modeling	[15]
Core computational framework for CITE-seq (HSC/LSC immunophenotyping)	totalVI	Multi-omics integration	Joint latent space (RNA + protein), batch correction, missing-protein handling → directly supports HSC/LSC surface phenotype + transcriptional state integration	[15]
C: Mechanistic “stemness regulators”: GRNs, TF programs, causality-leaning inference
In silico TF perturbation to test “does this TF maintain HSC identity or drive differentiation?”	CellOracle	GRN inference + perturbation simulation	Demonstrated in mouse and human hematopoiesis; supports mechanistic framing in a stem-cell review	[14] and GitHub[96]
TF-activity inference from scRNA-seq to link stage-specific regulators with differentiation states	SCENIC	TF network inference	Regulatory network + regulon activity per cell; can be used to interpret HSC/LSC programs	[13]
D: TI (to resolve rare HSC states, branching, cycling)
Capturing complex hematopoietic branching while preserving rare populations (useful for “rare HSC/pre-LSC” arguments)	VIA	TI	Lazy-teleporting random walks + MCMC; marketed as scalable/generalized TI	[5]
E: Handling missing modalities + batch effects in HSC multi omics (practical translational reality)
Multi-omic integration when HSC datasets are incomplete/unpaired across assays	scMaui	Multi-omics integration	Product-of-experts VAE; batch + missing modality handling	[16]
F: Aging/activation phenotypes in HSCs (imaging + AI)
Imaging-based “biological age” predictor from HSC nuclear architecture	ChromAgeNet	Chromatin aging prediction	CNN on 3D DAPI chromatin images; positions aging as a learnable HSC phenotype	[6]
G: LSC persistence proxies in patients: MRD/residual disease
Interpretable ML that supports MRD assessment (detecting residual immature/LSC-like compartments)	MAGIC-DR	MRD detection	Interpretable ML-guided approach for AML MRD	PubMed landing[56]
Notable for making a large AML flow cytometry dataset publicly available for benchmarking computational MRD tools	Computational MRD assessment (GMM + novelty detection)	MRD detection	Standardization + automated MRD explicitly discusses heterogeneity issues that connect to LSC persistence framing	[97]

¹Dataset: Cord blood hematopoietic stem cell single-cell RNA sequencing data (CD34+ vs CD34- populations) publicly available at NCBI GEO:GSE237832 (hyperlink: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE237832).

HSC: Hematopoietic stem cell; LSC: Leukemic stem cell; AML: Acute myeloid leukemia; scRNA-seq: Single-cell RNA sequencing; HSPC: Hematopoietic stem and progenitor cell; TI: Trajectory inference; GEO: Gene Expression Omnibus; EGA: European Genome-phenome Archive; TF: Transcription factor; scATAC-seq: Single-cell assay for transposase-accessible chromatin sequencing; PBMC: Peripheral blood mononuclear cell; CITE-seq: Cellular indexing of transcriptomes and epitopes by sequencing; GRN: Gene regulatory network; SCENIC: Single-cell regulatory network inference and clustering; VIA: Voyager in single-cell RNA sequencing analysis; MCMC: Markov chain Monte Carlo; VAE: Variational autoencoder; AI: Artificial intelligence; CNN: Convolutional neural network; DAPI: 4’,6-diamidino-2-phenylindole; MRD: Measurable residual disease; MAGIC-DR: Machine-learning guided approach to acute myeloid leukemia measurable residual disease detection in residual disease; ML: Machine learning; GMM: Gaussian mixture model.

Table 3 Summary of key machine learning algorithms, applications, and trade-offs in hematopoietic stem cell research

Machine learning algorithm	Primary HSC/hematology applications	Key advantages (Pros)	Key limitations (Cons)	Ref.
CNN	Morphological classification of HSCs vs MPPs; 3D chromatin age prediction (ChromAgeNet); automated quality control imaging in biomanufacturing	Unparalleled performance on spatial and image data; extracts features autonomously without requiring manual gating or human-defined parameters	Highly opaque “black box” nature requiring XAI for interpretability; demands massive, accurately annotated image datasets to train	[6,11,34]
Random forest/ensemble trees	Flow cytometric LSC phenotyping; early relapse detection; predicting cord blood CD34+ cell yield	Highly robust to overfitting; handles tabular clinical and multi-omics data effectively; naturally provides feature importance rankings	Less effective than deep learning for highly unstructured data (like raw images or free text); can struggle with extrapolating data outside the training range	[22,35]
Gradient Boosting (e.g., XGBoost)	Automated MRD detection in flow cytometry (MAGIC-DR); predicting synergistic drug combinations; venetoclax response prediction	Exceptional predictive accuracy on structured clinical and omics data; handles missing data well; highly scalable for large patient cohorts	Prone to overfitting on very small sample sizes; hyperparameter tuning is complex and computationally expensive	[23,56]
Deep learning/ANN	Multi-omics integration (e.g., totalVI); modeling age-dependent HSC self-renewal; multi-center AML survival prediction	Can capture extremely complex, non-linear biological relationships across massive, high-dimensional datasets (e.g., integrating RNA and protein expression)	Computationally intensive; high risk of learning artifactual batch effects rather than true biology if data is not strictly harmonized	[10,15,17]
NLP	Extracting HLA genotypes from unstructured electronic health records; predicting and phenotyping acute and chronic GVHD from clinical notes	Unlocks vast amounts of unstructured, historical clinical data that is otherwise inaccessible to standard statistical models	Performance is heavily dependent on the quality, consistency, and language of physician documentation; it struggles with implicit clinical context	[40-42]
RL	Dynamic optimization of ex vivo HSC expansion; adaptive control of bioreactor parameters (cytokines, perfusion, metabolic flux)	Enables continuous, autonomous, real-time process optimization without requiring a pre-defined static protocol	Requires highly accurate “digital twins” or simulation environments to train the agent safely; initial validation in GMP environments is regulatory complex	[37,77,78]

HSC: Hematopoietic stem cell; CNN: Convolutional neural network; MPP: Multipotent progenitor; XAI: Explainable artificial intelligence; LSC: Leukemic stem cell; XGBoost: Extreme Gradient Boosting; MRD: Measurable residual disease; MAGIC-DR: Machine-learning guided approach to acute myeloid leukemia measurable residual disease detection in residual disease; ANN: Artificial neural network; AML: Acute myeloid leukemia; NLP: Natural language processing; HLA: Human leukocyte antigen; GVHD: Graft-vs-host disease; RL: Reinforcement learning; GMP: Good Manufacturing Practice.

Table 4 Artificial intelligence performance vs conventional methods in acute myeloid leukemia and multiple myeloma risk stratification

Disease	Conventional method	AI method	Key performance advantages	Ref.
AML	ELN risk stratification	Multi-omics deep learning	Outperforms ELN-based approaches. Identifies LSC burden (not quantified by ELN). > 90% accuracy in therapy resistance prediction. Integrates clinical, cytogenetic, and molecular data	[53]
AML	Traditional cytogenetic/molecular classification	Supervised machine learning	Superior prediction of complete remission and 2-year survival. External validation confirms generalizability	[17]
MM	Standard risk assessment	Multi-omics integration (neural networks)	Accuracy in therapy resistance prediction. Enhanced drug resistance prediction. Integration of genomic biomarkers and clinical parameters	[31,32,98]
MM	Standard histopathology	Deep learning (MoSaicNet & AwareNet)	Spatial heterogeneity detection beyond cell density. Differentiates MGUS from MM based on spatial architecture	[30]
HSCT GVHD	Cox proportional hazard models	CNN-NLP hybrid	Superior risk stratification. Stratifies aGVHD incidence from 31.8% to 54.8%. Processes detailed HLA information vs binary matching	[41]
HSCT cGVHD	NIH consensus severity criteria	Machine learning phenotyping	Identified 7 distinct phenotypes. 2.24-fold mortality difference between risk groups. Better survival stratification than traditional scores	[42]

AI: Artificial intelligence; AML: Acute myeloid leukemia; ELN: European Leukemia Net; LSC: Leukemic stem cell; MM: Multiple myeloma; MGUS: Monoclonal gammopathy of undetermined significance; HSCT: Hematopoietic stem cell transplantation; GVHD: Graft-vs-host disease; CNN: Convolutional neural network; NLP: Natural language processing; aGVHD: Acute graft-vs-host disease; HLA: Human leukocyte antigen; cGVHD: Chronic graft-vs-host disease; NIH: National Institutes of Health.

Table 5 Validation checklist for leukemic stem cell/hematopoietic stem cell artificial intelligence models, adapted from TRIPOD + AI and PROBAST + AI guidelines with hematopoietic stem cell/Leukemic stem cell-specific requirements

Validation domain	Key requirement	LSC/HSC-specific considerations
Cohort representativeness	The training cohort must represent the target clinical population with respect to age, disease stage, and treatment era	LSC/HSC models should include balanced representation of ELN risk categories, stem-cell compartment measurements (CD34+CD38- frequencies), and both newly diagnosed and relapsed/refractory patients[52,106-108]
Event counts	An adequate number of outcome events (relapse, death, MRD positivity) to prevent overfitting	Minimum 10-20 events per predictor variable; for LSC-specific endpoints (e.g., LSC+ vs LSC-), ensure sufficient LSC+ cases across validation sets[52,55]
Internal validation	Model performance assessed on held-out data from the same source (cross-validation or hold-out split)	Report performance metrics (AUROC, calibration) separately for LSC-enriched vs LSC-depleted subgroups if the model claims to encode stemness biology[52,109,110]
External validation	Independent cohort from a different institution, time period, or geography	Essential for LSC models given center-to-center variability in LSC phenotyping protocols and MRD detection thresholds[52,106,110]
Prospective evaluation	Forward-looking validation on newly enrolled patients before clinical deployment	Required for LSC-targeted therapy selection models to confirm that AI predictions align with clinical outcomes under prospective conditions[55,109,111]
Dataset shift detection	Assess whether model performance degrades when applied to data with distributional differences (batch effects, assay drift)	Critical for flow cytometry-based LSC models: Validate across different antibody panels, fluorophores, and cytometers; report performance stratified by batch[55,112]
Calibration assessment	Predicted probabilities should match observed event frequencies	For LSC burden models, calibration plots should show agreement between predicted LSC frequency (or surrogate score) and directly measured LSC% by flow cytometry in the calibration subset[52,55,106]
Decision-curve analysis	Net benefit of model-guided decisions compared to treat-all or treat-none strategies	For LSC-directed therapies (venetoclax, Menin inhibitors), decision curves should quantify clinical utility across risk thresholds relevant to treatment intensification decisions[55,111]
Explainability and feature attribution	Use of XAI methods (SHAP, attention weights) to identify which features drive predictions	Essential for validating LSC-AI hypothesis: Determine whether high-risk predictions are driven by known stemness genes (17-gene LSC score, HOXMEIS1 programs) or alternative pathways[52,113]
Bias and fairness evaluation	Assess performance stratified by demographic subgroups and underrepresented populations	Evaluate whether LSC models perform equivalently across age groups (pediatric vs adult vs elderly AML), ancestry, and sex; report subgroup-specific metrics[107]
Missing data handling	Transparent reporting of missingness patterns and imputation strategies	LSC models often integrate multi-omics data with heterogeneous completeness (e.g., scRNA-seq available for a subset); clearly document handling of missing modalities and proteins in CITE-seq[111]
Comparator benchmarking	Performance compared to established clinical risk systems	For AML LSC models, benchmark against ELN 2022 risk classification, 17-gene LSC score, and LSC frequency by flow cytometry; report incremental predictive value[114]

LSC: Leukemic stem cell; HSC: Hematopoietic stem cell; ELN: European Leukemia Net; MRD: Measurable residual disease; AUROC: Area under the receiver operating characteristic; AI: Artificial intelligence; XAI: Explainable artificial intelligence; SHAP: SHapley Additive exPlanations; AML: Acute myeloid leukemia; scRNA-seq: Single-cell RNA sequencing; CITE-seq: Cellular indexing of transcriptomes and epitopes by sequencing.