1 of 64

Synthetic Data Generation, Multi-Dimensional

Validation, and Clinically Actionable Risk

Stratification in a FIL-DLBCL Clinical Cohort

2 of 64

The Clinical Context

Diffuse Large B-Cell Lymphoma is the most common aggressive non-Hodgkin lymphoma, whose incidence increases with age

"Large B cell lymphoma – cytology" by Nephron, Wikimedia Commons, CC BY-SA 3.0

3 of 64

The Clinical Context

Diffuse Large B-Cell Lymphoma is the most common aggressive non-Hodgkin lymphoma, whose incidence increases with age

"Large B cell lymphoma – cytology" by Nephron, Wikimedia Commons, CC BY-SA 3.0

For elderly patients, FIL defined the Elderly Prognostic Index (EPI): a validated 3-tier risk score built from clinical, disease and geriatric variables, using the data from the Elderly Cohort

4 of 64

The Clinical Context

Diffuse Large B-Cell Lymphoma is the most common aggressive non-Hodgkin lymphoma, whose incidence increases with age

For elderly patients, FIL defined the Elderly Prognostic Index (EPI): a validated 3-tier risk score built from clinical, disease and geriatric variables, using the data from the Elderly Cohort

DLBCL patients aged > 65

1065 patients, 19 mixed type features

EPI components: sCGA + IPI + hemoglobin

Outcome variable: Fail

EPI risk score: Low, Intermediate, High

​

5 of 64

The Research Goal

Fixed scoring rules can struggle to capture deep multivariate relationships and generalize to new cohorts

6 of 64

The Research Goal

Fixed scoring rules can struggle to capture deep multivariate relationships and generalize to new cohorts

EPI gives a 3-tier stratification rule, but some patients (especially high risk) remain discordant with outcome

7 of 64

The Research Goal

Fixed scoring rules can struggle to capture deep multivariate relationships and generalize to new cohorts

EPI gives a 3-tier stratification rule, but some patients (especially high risk) remain discordant with outcome

DISCORDANT SUBGROUP

8 of 64

Core Objective: Can we actively refine a validated clinical rule like the EPI using generative AI?

The Research Goal

Fixed scoring rules can struggle to capture deep multivariate relationships and generalize to new cohorts

EPI gives a 3-tier stratification rule, but some patients (especially high risk) remain discordant with outcome

DISCORDANT SUBGROUP

9 of 64

Synthetic Data

Core Objective: Can we actively refine a validated clinical rule like the EPI using generative AI?

10 of 64

Training and benchmarking models without privacy or size concerns

Synthetic Data

Core Objective: Can we actively refine a validated clinical rule like the EPI using generative AI?

11 of 64

Training and benchmarking models without privacy or size concerns

Synthetic Data

Augmenting real data (rare cases, imbalances)

Core Objective: Can we actively refine a validated clinical rule like the EPI using generative AI?

12 of 64

Training and benchmarking models without privacy or size concerns

Synthetic Data

Augmenting real data (rare cases, imbalances)

Core Objective: Can we actively refine a validated clinical rule like the EPI using generative AI?

Creating synthetic control arms for more efficient resource management

13 of 64

Dual Framework

GENERATIVE

Per-split training

generative models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic data evaluation

Fidelity • Privacy • Utility

14 of 64

Dual Framework

GENERATIVE

Per-split training

generative models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic data evaluation

Fidelity • Privacy • Utility

Standard statistical evaluation might not be enough!

15 of 64

Dual Framework

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

GENERATIVE

Per-split training

generative models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic data evaluation

Fidelity • Privacy • Utility

16 of 64

Dual Actionability Framework

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

SHARED BACKBONE

10 independent�stratified splits

1,065 patients

full cohort

85 / 15 stratified split

repeated independently

905

TRAIN

160

LOCKED REAL TEST

SAME SPLIT → SAME TEST SET

GENERATIVE

Per-split training

generative models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

The same independent train/test partition is preserved from generation through downstream clinical evaluation.

17 of 64

Dual Actionability Framework

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

SHARED BACKBONE

10 independent�stratified splits

1,065 patients

full cohort

85 / 15 stratified split

repeated independently

905

TRAIN

160

LOCKED REAL TEST

SAME SPLIT → SAME TEST SET

GENERATIVE

Per-split training

generative models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

The same independent train/test partition is preserved from generation through downstream clinical evaluation.

18 of 64

Pool of generative models

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

19 of 64

Pool of generative models

State of the Art

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

20 of 64

Pool of generative models

CTAB-GAN

Adversarial network specifically tailored with encodings for mixed-type variables and class imbalance

State of the Art

Zhao et al., ACML 2021

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

21 of 64

Pool of generative models

CTAB-GAN

Adversarial network specifically tailored with encodings for mixed-type variables and class imbalance

Tabsyn

Diffusion model achieving high statistical fidelity by iteratively reversing a noise corruption process

State of the Art

Zhao et al., ACML 2021

Zhang et al., ICLR 2024

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

22 of 64

Pool of generative models

CTAB-GAN

Adversarial network specifically tailored with encodings for mixed-type variables and class imbalance

Tabsyn

Diffusion model achieving high statistical fidelity by iteratively reversing a noise corruption process

REaLTabFormer (RTF)

Autoregressive transformer that models tabular feature dependencies sequentially, token by token

State of the Art

Zhao et al., ACML 2021

Zhang et al., ICLR 2024

Solatorio et al., arXiv 2023

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

23 of 64

Pool of generative models

𝛽-VAE

Learns a disentangled continuous latent space to isolate independent clinical risk factors.

CTAB-GAN

Adversarial network specifically tailored with encodings for mixed-type variables and class imbalance

Tabsyn

Diffusion model achieving high statistical fidelity by iteratively reversing a noise corruption process

REaLTabFormer (RTF)

Autoregressive transformer that models tabular feature dependencies sequentially, token by token

Proposed/Reimplemented

State of the Art

Zhao et al., ACML 2021

Zhang et al., ICLR 2024

Solatorio et al., arXiv 2023

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

24 of 64

Pool of generative models

𝛽-VAE

Learns a disentangled continuous latent space to isolate independent clinical risk factors.

CTAB-GAN

Adversarial network specifically tailored with encodings for mixed-type variables and class imbalance

Literature-Informed LLM

Fine-tuned on DLBCL literature; generates coherent patient records directly from semantic priors

Tabsyn

Diffusion model achieving high statistical fidelity by iteratively reversing a noise corruption process

REaLTabFormer (RTF)

Autoregressive transformer that models tabular feature dependencies sequentially, token by token

Proposed/Reimplemented

State of the Art

Zhao et al., ACML 2021

Zhang et al., ICLR 2024

Solatorio et al., arXiv 2023

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

25 of 64

Pool of generative models

𝛽-VAE

Learns a disentangled continuous latent space to isolate independent clinical risk factors.

CTAB-GAN

Adversarial network specifically tailored with encodings for mixed-type variables and class imbalance

Literature-Informed LLM

Fine-tuned on DLBCL literature; generates coherent patient records directly from semantic priors

Tabsyn

Diffusion model achieving high statistical fidelity by iteratively reversing a noise corruption process

Transformer Flow Matching (TFM)

Couples optimal transport flow matching with a tokenized transformer to jointly model mixed-type clinical data

REaLTabFormer (RTF)

Autoregressive transformer that models tabular feature dependencies sequentially, token by token

Proposed/Reimplemented

State of the Art

Zhao et al., ACML 2021

Zhang et al., ICLR 2024

Solatorio et al., arXiv 2023

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

26 of 64

Transformer Flow Matching

Transformer Flow Matching (TFM)

Couples optimal transport flow matching with a tokenized transformer to jointly model mixed-type clinical data

Input mixed data

d-model space

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

First, we represent the mixed clinical table as feature tokens, leading to a commont representation of continuous and categorical variables

27 of 64

Transformer Flow Matching

Transformer Flow Matching (TFM)

Couples optimal transport flow matching with a tokenized transformer to jointly model mixed-type clinical data

Input mixed data

d-model space

train vector field

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

Then, we use optimal-transport conditional flow matching.

Conceptually, the model learns a vector field that transports noise toward the empirical clinical distribution.

28 of 64

Transformer Flow Matching

Transformer Flow Matching (TFM)

Couples optimal transport flow matching with a tokenized transformer to jointly model mixed-type clinical data

Input mixed data

d-model space

generate

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

train vector field

Once that vector field has been learned, we can start from noise and follow the learned flow to generate new patients.

29 of 64

Generative Framework - Results

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

The colour scale gives a quick visual indication of performance: the stronger the colour, the better the value for that metric, so it lets us compare models across each row at a glance

30 of 64

Generative Framework - Results

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

Each row corresponds to a different aspect of synthetic-data quality, moving from fidelity, to privacy, to downstream utility

31 of 64

Generative Framework - Results

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

We are looking for a balanced profile across the three dimensions

TFM stands out: it remains consistently strong across fidelity, privacy and utility

​

32 of 64

Generative Framework - Results

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

No single model is the clear winner

33 of 64

Dual Framework

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

SHARED BACKBONE

10 independent�stratified splits

1,065 patients

full cohort

85 / 15 stratified split

repeated independently

905

TRAIN

160

LOCKED REAL TEST

SAME SPLIT → SAME TEST SET

GENERATIVE

Per-split training

generatuve models from our proposed + sota pool

Synthetic cohort

5,000 records / split�inverse-transformed

Synthetic-data evaluation

Fidelity • Privacy • Utility

REPEAT ×10

The same independent train/test partition is preserved from generation through downstream clinical evaluation.

34 of 64

Clinical Actionability Framework

CLINICAL

ACTIONABILITY

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REAL TRAIN

905 patients

AutoML-Med

selects the downstream�binary classifier

SELECTED MODEL

Logistic regression

01

Freeze the downstream pipeline on REAL training data

35 of 64

Clinical Actionability Framework

REAL BASELINE

905 REAL TRAIN

TSTR

5,000 SYNTHETIC

AUGMENTATION

905 REAL TRAIN

+

5,000 SYNTHETIC

CLINICAL

ACTIONABILITY

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

02

Three training scenarios for the frozen classifier

36 of 64

From one mortality probability to the clinical metrics

03

Clinical Actionability Framework

The frozen binary classifier returns one continuous risk score; the same score drives all downstream analyses

P(death|X)

continuous risk score

CONTINUOUS SCORE

C-index · IDI · NRI · Net Benefit

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

37 of 64

From one mortality probability to the clinical metrics

03

Clinical Actionability Framework

The frozen binary classifier returns one continuous risk score; the same score drives all downstream analyses

P(death|X)

continuous risk score

CONTINUOUS SCORE

C-index · IDI · NRI · Net Benefit

TRAIN-DERIVED CUTPOINTS

33rd / 66th percentiles

​

Log-rank

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

38 of 64

Clinical Actionability - Metrics

Six complementary metrics to capture discrimination, survival separation, reclassification and clinical decision value

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

39 of 64

Results

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

40 of 64

Results

TFM TSTR preserves baseline discrimination

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

41 of 64

Results

TFM TSTR preserves baseline discrimination

​

Aug-TFM gives the highest C-index, but gains are significant in IDI/NRI/net benefit

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

42 of 64

Results

TFM TSTR preserves baseline discrimination

​

Aug-TFM gives the highest C-index, but gains are significant in IDI/NRI/net benefit

​

LLM shows the largest reclassification & net-benefit gains

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

43 of 64

Results

TFM TSTR preserves baseline discrimination

​

Aug-TFM gives the highest C-index, but gains are significant in IDI/NRI/net benefit

​

LLM shows the largest reclassification & net-benefit gains

​

β-VAE underperforms on downstream discrimination

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

44 of 64

Results

TFM TSTR preserves baseline discrimination

​

Aug-TFM gives the highest C-index, but gains are significant in IDI/NRI/net benefit

​

LLM shows the largest reclassification & net-benefit gains

​

β-VAE underperforms on downstream discrimination

​

CTAB-GAN: highest variability, lowest discrimination, consistent with mode-collapse effects

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

45 of 64

Clinical Actionability Framework

P(death|X)

continuous risk score

CONTINUOUS SCORE

C-index · IDI · NRI · Net Benefit

TRAIN-DERIVED CUTPOINTS

33rd / 66th percentiles

​

Log-rank

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

46 of 64

Clinical Actionability Framework

P(death|X)

continuous risk score

CONTINUOUS SCORE

C-index · IDI · NRI · Net Benefit

TRAIN-DERIVED CUTPOINTS

33rd / 66th percentiles

​

Log-rank

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

SHAP ANALYSIS

Feature importance for new SHAP-informed EPI definition

From SHAP explanations to a refined prognostic score

04

47 of 64

The discordant subroup: 14.5% of the cohort (n=154) are patients assigned to the high-risk class but survive

From SHAP explanations to a refined prognostic score

​

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

STRONGEST RISK SIGNALS

sCGA · age

push the model toward death (geriatric frailty)

48 of 64

The discordant subroup: 14.5% of the cohort (n=154) are patients assigned to the high-risk class but survive

From SHAP explanations to a refined prognostic score

​

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

STRONGEST RISK SIGNALS

sCGA · age

push the model toward death (geriatric frailty)

MISSING FROM THE EPI

albumin · IADL

protective information not used for base epi

49 of 64

From SHAP explanations to a refined prognostic score

04

Clinical Actionability Framework

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

SHAP ANALYSIS

Feature importance for new SHAP-informed EPI definition

Explain the classifier

Compute feature contributions within the discordant subgroup

50 of 64

From SHAP explanations to a refined prognostic score

04

Clinical Actionability Framework

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

SHAP ANALYSIS

Feature importance for new SHAP-informed EPI definition

Explain the classifier

Compute feature contributions within the discordant subgroup

Keep selected features

with mean |SHAP|≥0.025

51 of 64

From SHAP explanations to a refined prognostic score

04

Clinical Actionability Framework

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

SHAP ANALYSIS

Feature importance for new SHAP-informed EPI definition

Explain the classifier

Compute feature contributions within the discordant subgroup

Keep selected features

with mean |SHAP|≥0.025

Fit Cox model

using all retained features for each scenario

52 of 64

From SHAP explanations to a refined prognostic score

04

Clinical Actionability Framework

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REPEAT ×10

SHAP ANALYSIS

Feature importance for new SHAP-informed EPI definition

Explain the classifier

Compute feature contributions within the discordant subgroup

Keep selected features

with mean |SHAP|≥0.025

Fit Cox model

using all retained features for each scenario

Create refined score

converting the cox risk predictors in a 3-tier risk score

53 of 64

The discordant subroup: 14.5% of the cohort (n=154) are patients assigned to the high-risk class but survive

From SHAP explanations to a refined prognostic score

​

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

sCGA · age

push the model toward death (geriatric frailty)

albumin · IADL

protective information not used for base epi

TFM SHAP-INFORMED COX SCORE

5 EPI FEATURES RETAINED

5 NON-EPI FEATURES ADDED

age · Ann Arbor stage · ECOG�extranodal sites · sCGA

albumin · IADL · B-symptoms�bulky disease · sex

LDH + haemoglobin (present in original EPI) were not selected by the TFM SHAP analysis

STRONGEST RISK SIGNALS

MISSING FROM THE EPI

54 of 64

The discordant subroup: 14.5% of the cohort (n=154) are patients assigned to the high-risk class but survive

From SHAP explanations to a refined prognostic score

​

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

sCGA · age

push the model toward death (geriatric frailty)

albumin · IADL

protective information not used for base epi

REFINED TEST-SET SCORE

C-index

0.679

→

0.731

+0.052

Log-rank

19.83

→

22.18

+2.35

TFM SHAP-INFORMED COX SCORE

5 EPI FEATURES RETAINED

5 NON-EPI FEATURES ADDED

age · Ann Arbor stage · ECOG�extranodal sites · sCGA

albumin · IADL · B-symptoms�bulky disease · sex

LDH + haemoglobin (present in original EPI) were not selected by the TFM SHAP analysis

STRONGEST RISK SIGNALS

MISSING FROM THE EPI

55 of 64

From SHAP explanations to a refined prognostic score

​

CLINICAL

ACTIONABILITY

Same TRAIN/TEST split

Frozen pipeline

AutoML-Med + preprocessing�selected on TRAIN only

Three training scenarios

Real baseline

TSTR�Augmentation

Evaluate on TEST

C-index • Log-rank • IDI • NRI�Net Benefit • SHAP

REFINED TEST-SET SCORE

C-index

0.679

→

0.731

+0.052

Log-rank

19.83

→

22.18

+2.35

TFM-defined EPI

TFM SHAP-INFORMED COX SCORE

5 EPI FEATURES RETAINED

5 NON-EPI FEATURES ADDED

age · Ann Arbor stage · ECOG�extranodal sites · sCGA

albumin · IADL · B-symptoms�bulky disease · sex

LDH + haemoglobin (present in original EPI) were not selected by the TFM SHAP analysis

56 of 64

  • The Analytical Target: 14.5% of the cohort (n=154) are patients assigned to the highest-risk class by the traditional EPI who ultimately survive
  • SHAP analysis reveals the EPI high-risk label is heavily driven by the geriatric frailty score (sCGA), penalizing patients for age-related frailty rather than disease aggressiveness

​

  • The AI Discovery: TFM captures predictive signals missed by the historical index, relying heavily on markers absent from the EPI: specifically serum albumin and preserved instrumental autonomy (IADL)
  • A Refined Clinical Score: An interpretable AI-driven Cox score built from these SHAP-identified features raises the C-index to 0.731 (up from the EPI's 0.678) and significantly sharpens risk separation and stratification

​

Explainability: The "Discordant Survivor" subgroup

57 of 64

58 of 64

Conclusions

01 SYNTHETIC DATA

Fidelity alone is not enough; augmentation is the meaningful downstream test for clinical actionability

59 of 64

Conclusions

01 SYNTHETIC DATA

Fidelity alone is not enough; augmentation is the meaningful downstream test for clinical actionability

02 TFM

TSTR preserves baseline; augmentation adds significant reclassification and net-benefit gains

60 of 64

Conclusions

01 SYNTHETIC DATA

Fidelity alone is not enough; augmentation is the meaningful downstream test for clinical actionability

02 TFM

TSTR preserves baseline; augmentation adds significant reclassification and net-benefit gains

03 REFINEMENT

SHAP reveals signals outside EPI and supports an interpretable refined score

61 of 64

Conclusions

01 SYNTHETIC DATA

Fidelity alone is not enough; augmentation is the meaningful downstream test

02 TFM

TSTR preserves baseline; augmentation adds significant reclassification and net-benefit gains

03 REFINEMENT

SHAP reveals signals outside EPI and supports an interpretable refined score

Take-home message:

High quality synthetic cohorts can not only preserve the signal needed by downstream models but also enrich it

62 of 64

Conclusions

01 SYNTHETIC DATA

Fidelity alone is not enough; augmentation is the meaningful downstream test for clinical actionability

02 TFM

TSTR preserves baseline; augmentation adds significant reclassification and net-benefit gains

03 REFINEMENT

SHAP reveals signals outside EPI and supports an interpretable refined score

FUTURE WORK

​

  • external validation
  • clinician assessment
  • longitudinal series generation

Take-home message:

High quality synthetic cohorts can not only preserve the signal needed by downstream models but also enrich it

63 of 64

Take home message:

  • Statistical Fidelity is often insufficient: Generative AI in healthcare must be evaluated on downstream clinical actionability, not just distributional similarity
  • TFM Solves the Mixed-Type Challenge: Transformer Flow Matching effectively models complex clinical data, achieving top-tier statistical fidelity while preserving privacy at lower computational cost
  • Synthetic Data Refines Prognosis: Augmenting real cohorts with TFM-generated records significantly improves discriminative power, clinical reclassification, and survival risk separation over historical baselines

Future Directions:

  • External Validation: Testing on independent international registries and extending to other indices like the MIPI and IPI
  • Clinical Translation: Prospective assessment of the SHAP-derived score directly with hematologists
  • Architectural Extension: Upgrading the TFM generator to handle native time-to-event and competing-risk structures for direct survival modeling

Conclusions

64 of 64