FINAL_IMPROVEMENT_RECOMMENDATIONS.md

Final Improvement Recommendations: Path to Top 5% Success Rate

Priority-Ranked Action Items for Grant Proposal Excellence

Document Purpose: Synthesize critical evaluation and peer review simulation into actionable, priority-ranked recommendations Target Outcome: Elevate proposal from current top 10-15% to top 3-5% tier Evaluation Baseline: Current composite score 7.8/9.0 → Target 8.5-9.0/9.0

---

EXECUTIVE SUMMARY: CRITICAL PATH TO TOP 5%

Current Assessment

• Composite Score: 7.8/9.0 (Top 10-15% currently)
• Success Probability: 65-75%
• Status: Fundable with major revisions, but NOT in top 5% tier

Projected Assessment (After Implementing Recommendations)

• Projected Composite Score: 8.5-9.0/9.0 (Top 3-5%)
• Projected Success Probability: 85-95%+
• Status: Highly competitive, likely funding with potential for exceptional rating

The 3 Critical Barriers to Top 5%

BARRIER #1: Investigator Credibility Gap (-1.5 to -2.0 points)

• Zero named investigators, zero track record, zero preliminary data
• Reviewers cannot assess feasibility for $50M, 7-year, 50-site study without PI credentials
• Fix: Name PI + Co-Is, add CVs, preliminary data, letters of support

BARRIER #2: INCITE Model Status Ambiguity (-1.0 to -1.5 points)

• NeuroX-Fusion 130B described as existing but not cited or validated
• Unclear if model is real (low risk) or must be built from scratch (high risk, +$10M, +12 months)
• Fix: Clarify model status, provide citation/preliminary results OR add pre-training aim

BARRIER #3: 50-Site Coordination Feasibility (-0.5 to -1.0 points)

• Logistics of recruiting/retaining 50 sites across 5 continents severely underestimated
• No site management plan, no budget breakdown, no attrition modeling
• Fix: Add detailed coordination plan, site recruitment strategy, governance structure

Timeline to Fix: 4-8 weeks (with dedicated effort)

ROI: Fixing these 3 barriers → +2.5 to +3.5 points → 7.8 → 10.3-11.3 (capped at 9.0) = 8.5-9.0 realistic score

---

PRIORITY 1: CRITICAL FIXES (MANDATORY FOR TOP 5%)

Definition: These fixes are non-negotiable. Without them, proposal cannot reach top 5% regardless of other strengths.

Estimated Impact: +2.0 to +2.5 points overall (7.8 → 9.8-10.3, realistically 8.5-9.0)

Timeline: 4-6 weeks with full team effort

---

1.1 NAME INVESTIGATORS WITH TRACK RECORDS [HIGHEST PRIORITY]

Current State: No investigators named, no CVs, no publication lists

Required Fix:

Step 1: Name Principal Investigator (PI)

Required PI Profile:

• Senior investigator (Professor level, 15+ years experience)
• Multi-site expertise: Led 5+ multi-site studies (ideally 10+ sites, 1,000+ participants)
• Autism/DD expertise: 50+ autism/DD publications, h-index ≥50
• Funding track record: $20M+ in prior NIH/NSF grants (10+ R01-equivalent awards)
• ADOS-2/neuroimaging credentials: Gold-standard diagnostic training OR multi-site neuroimaging leadership

Example PI Profile (Hypothetical):

Dr. Jane Smith, PhD Professor of Psychiatry and Neuroscience, University of [X] Director, Center for Autism Research Excellence

Track Record:

• 25 years autism research experience
• Principal Investigator, ENIGMA-Autism Working Group (40 sites, 15,000+ participants)
• Co-Investigator, ABIDE consortium (multiple sites, 1,200+ participants)
• 180 peer-reviewed publications (h-index 85, 25,000+ citations)
• $35M in NIH/NSF funding over 15 years (12 R01s, 3 P50 center grants)
• ADOS-2 certified evaluator (2005), trainer (2010)
• 15+ PhD students graduated (12 now faculty at R1 universities)

Preliminary Work Relevant to This Proposal:

• Pilot multimodal fusion study (n=100, AUC 0.88) - manuscript in preparation
• Federated learning simulation on ABIDE (89% inter-site accuracy) - presented at OHBM 2024
• INCITE compute allocation secured (3M core-hours on Aurora, 2025-2026)

Step 2: Name 4-6 Co-Investigators

Required Expertise Coverage (at minimum):

Co-I #1: AI/ML Foundation Model Expert

• Requirements:
  • Foundation model development experience (ideally 10B+ parameter models)
  • Federated learning publications (5+ papers in NeurIPS, ICML, ICLR)
  • Medical AI deployment experience (bonus: clinical ML systems)
• Example: Former Google Brain/Meta AI/OpenAI researcher OR academic with h-index ≥40 in ML

Co-I #2: Child Psychiatrist / Clinical Trials Expert

• Requirements:
  • ADOS-2/ADI-R certified evaluator (essential)
  • 10+ autism clinical trials as PI or site PI
  • Pragmatic trial experience (pRCT design, real-world effectiveness)
• Example: Academic child psychiatrist with 100+ autism patients evaluated, 20+ RCTs

Co-I #3: Genetic Epidemiologist / Genomics Expert

• Requirements:
  • WES analysis expertise (GATK pipeline, rare variant burden tests)
  • GWAS experience (preferably autism GWAS, e.g., Grove et al. 2019 co-author)
  • Causal inference in genomics (Mendelian randomization)
• Example: Investigator with 50+ genetics papers, h-index ≥30

Co-I #4: Neuroimaging Expert / Multi-Site Coordination

• Requirements:
  • Multi-site neuroimaging leadership (preferably ENIGMA, ABIDE, or equivalent)
  • FreeSurfer, fMRI preprocessing, quality control expertise
  • Harmonization experience (ComBat, traveling phantom, etc.)
• Example: ENIGMA working group leader OR ABIDE contributor with 30+ neuroimaging papers

Co-I #5: Regulatory Scientist / FDA Consultant

• Requirements:
  • FDA De Novo submission experience (5+ successful submissions)
  • Former FDA reviewer (bonus) OR regulatory affairs VP at medical device company
  • AI/ML SaMD expertise (21st Century Cures Act, FDA AI/ML guidance 2024)
• Example: Regulatory consultant with 10+ SaMD approvals, ISO 13485 QMS expertise

Co-I #6 (Optional): Biostatistician / Adaptive Trial Expert

• Requirements:
  • Bayesian adaptive trial design (interim analyses, futility/efficacy stopping)
  • Multi-site cluster randomization expertise
  • Federated learning statistics (differential privacy, site heterogeneity modeling)
• Example: Biostatistician with 50+ clinical trial papers, DSMB membership experience

Step 3: Provide CVs and Biosketches

For Each Investigator:

• NIH Biosketch (5 pages): Education, positions, honors, contributions to science, research support
• Key Publications (15 most relevant, with YOUR role highlighted)
• Prior Funding (Active and completed grants, total $$, your role)
• Preliminary Work (section in biosketch OR separate 2-3 page document)

Step 4: Add Preliminary Data

Minimum Viable Preliminary Data:

Option A: Pilot Multimodal Fusion Study (n=50-100)

• Sample: 50 ASD, 50 TD controls (from single site)
• Modalities: sMRI + fMRI (minimum), ideally + EEG/genomics
• Analysis: Multimodal fusion (early/intermediate/late fusion comparison)
• Results: AUC 0.85-0.90 (proof-of-concept that multimodal beats unimodal)
• Status: "Manuscript in preparation" OR "Presented at INSAR 2024"

Option B: Federated Learning Simulation (ABIDE Dataset)

• Sample: ABIDE dataset (1,112 participants, 17 sites)
• Method: Simulate federated learning (site-by-site training, FedAvg aggregation)
• Analysis: Leave-one-site-out cross-validation (17-fold)
• Results: Inter-site accuracy 85-90% (demonstrates FL feasibility)
• Status: "Presented at OHBM 2024" OR "Submitted to NeuroImage"

Option C: LoRA Fine-Tuning on Existing Foundation Model (BrainLM)

• Sample: n=50 DD patients (ASD or ADHD)
• Method: Fine-tune BrainLM (or BrainOmni for EEG) with LoRA (rank=8)
• Analysis: Compare LoRA (r=8) vs. full fine-tuning vs. zero-shot
• Results: LoRA achieves 95% of full fine-tuning performance with 1% parameters
• Status: "Manuscript in preparation"

Gold Standard Preliminary Data (if time/resources allow):

• All 3 of the above (multimodal pilot, FL simulation, LoRA fine-tuning)
• Plus: INCITE allocation secured + preliminary NeuroX-Fusion 130B results (even on non-DD task)
• Plus: 5-10 site letters of intent (for multi-site recruitment feasibility)

Step 5: Add Letters of Support

Required Letters (Minimum):

Letter #1: INCITE Program Director / Aurora Compute Allocation

• From: DOE INCITE program director OR Aurora supercomputer allocation manager
• Content:
  • Confirms compute allocation (e.g., "3M core-hours on Aurora for 2025-2026")
  • Supports scientific merit of NeuroX-Fusion 130B pre-training
  • States timeline (e.g., "Pre-training expected to complete Q2 2026")
• Critical: Without this letter, reviewers will assume INCITE model doesn't exist

Letter #2-6: Site Commitment Letters (5 sites minimum)

• From: Site PIs at 5 diverse locations (e.g., US academic, EU academic, Asia community clinic, US rural, Latin America)
• Content:
  • Commits to participate (recruit 60 participants over 5 years)
  • States IRB approval timeline (e.g., "IRB approval expected within 6 months of funding")
  • Confirms site capabilities (MRI scanner, EEG lab, genomics partnership, ADOS-2 certified staff)
  • Requests co-authorship, site-specific analyses, $100K site funding
• Impact: Demonstrates feasibility of 50-site recruitment (if 5 sites already committed, scaling to 50 is credible)

Letter #3-4: Advisory Board (2-3 senior leaders)

• From: Renowned autism researchers NOT on investigator team (e.g., SFARI director, IACC member, INSAR president)
• Content:
  • Endorses scientific approach ("This multimodal federated learning approach is innovative and timely")
  • Confirms unmet need ("Early diagnosis and precision medicine for autism are critical gaps")
  • States willingness to serve on Scientific Advisory Board (quarterly meetings)
• Impact: Signals field endorsement (reviewers trust advisory board judgment)

Letter #5: FDA Regulatory Consultant

• From: Regulatory expert (former FDA reviewer OR consultant with 5+ SaMD approvals)
• Content:
  • Confirms De Novo pathway feasibility ("Based on Canvas Dx precedent, De Novo is appropriate")
  • Estimates timeline to FDA clearance ("7-10 years realistic with robust pRCT validation")
  • States willingness to consult on regulatory strategy (pre-submission meetings, IDE/De Novo submissions)
• Impact: De-risks regulatory pathway (reviewers see expert guidance)

Letter #6 (Optional): Patient Advocacy Organization

• From: Autism Self-Advocacy Network (ASAN), Autistic Women & Nonbinary Network, or similar
• Content:
  • Endorses patient-centered outcomes (early diagnosis reduces family stress)
  • Confirms advisory role (autistic adults on study design team)
  • Raises ethical considerations (risk of labeling, need for genetic counseling)
• Impact: Shows community buy-in, addresses ethical concerns proactively

---

Deliverables for Fix 1.1:

• Named PI + Co-Is (6 investigators total)
• CVs/Biosketches (5 pages each, NIH format)
• Preliminary data report (5-10 pages: pilot study, FL simulation, OR LoRA results)
• Letters of support (6 minimum: INCITE, 5 sites, 2 advisory, 1 regulatory)

Estimated Effort: 4-6 weeks (if investigators already identified and willing)

Estimated Cost: $0 (if investigators volunteer) to $50K (if preliminary data collection needed)

Impact on Score:

• Investigators dimension: 7.2 → 8.5-9.0 (+1.3 to +1.8 points)
• Overall composite: 7.8 → 8.2-8.5 (+0.4 to +0.7 points)

---

1.2 CLARIFY INCITE NEUROX-FUSION 130B MODEL STATUS [CRITICAL]

Current State: Model described in detail but not cited, no preliminary results, unclear if exists

Required Fix:

Decision Point: Does NeuroX-Fusion 130B Exist?

Option A: Model Exists (Lower Risk, Preferred)

Action Items:

1. Cite the model:

   • Add citation: Paper, technical report, ArXiv preprint, OR DOE INCITE program website
   • Example: "NeuroX-Fusion 130B (Smith et al., 2025, ArXiv:2501.XXXXX)"
   • If no public paper yet: "NeuroX-Fusion 130B (INCITE 2025, https://www.alcf.anl.gov/incite/...)"

2. Provide preliminary results:

   • Show performance on any task (even general neuroscience, non-DD)
   • Examples:
     • "NeuroX-Fusion 130B achieves 0.92 AUC on BrainLM benchmark (predicting age, sex, cognitive scores from fMRI)"
     • "Zero-shot transfer to ABIDE: 78% accuracy (vs. BrainLM 75%, random 50%)"
   • Include 1-2 figures: Performance vs. model size, performance vs. training data size

3. Confirm access:

   • Attach INCITE allocation letter (see Fix 1.1, Letter #1)
   • State license: "Open-source under MIT license" OR "Proprietary, but we have license to fine-tune and commercialize"

4. Add architecture details:

   • Currently vague: "SwiFT 4D + Channel-equivariant + BrainOmni" - how are these integrated?
   • Add architecture diagram (1 page): Show how 3 sub-models are combined (ensemble? modular? hybrid?)
   • Clarify parameter breakdown: SwiFT (15B) + Channel-eq (30B) + BrainOmni (85B) = 130B (additive?) OR 130B total (shared parameters?)

Option B: Model Doesn't Exist (Higher Risk, Must Build)

Action Items:

1. Add Specific Aim 1: "Pre-train NeuroX-Fusion 130B Foundation Model"

   • Move this from "Background" to explicit research aim
   • Timeline: 12-18 months (Year 1-2)
   • Milestones:
     • Month 1-6: Data curation (ABIDE, ADHD-200, NDAR, HCP - total 50,000+ scans)
     • Month 7-12: Model training on Aurora (100 epochs, 10-15 days compute)
     • Month 13-18: Validation on held-out datasets, zero-shot transfer tests

2. Add budget for pre-training:

   • Compute: $5-10M (Aurora allocation OR cloud TPUs if INCITE unavailable)
   • Data licensing: $1-2M (NDAR, HCP data use agreements)
   • Personnel: $1M (2 ML engineers × 2 years × $250K each)
   • Total: $7-13M (add to $50M total → $57-63M)

3. Add fallback plan:

   • If Aurora unavailable or delayed: Use Google Cloud TPU v5 (estimate $3-5M for 130B model training)
   • If 130B infeasible: Fall back to BrainLM (3,662 subjects, 3B parameters, existing, open-source)
   • If compute budget insufficient: Reduce model size to 13B (10× smaller, 10× faster, $500K-1M compute)

4. Add risk mitigation:

   • Risk: 130B model training fails (technical issues, compute unavailable, insufficient data)
   • Mitigation: Start with smaller model (13B) in Year 1, scale to 130B in Year 2 if successful
   • Contingency: BrainLM (existing) ensures project can proceed even without custom foundation model

Recommendation: If NeuroX-Fusion 130B doesn't exist, seriously consider using BrainLM (existing) + LoRA fine-tuning instead of building 130B from scratch. This reduces risk, timeline, and cost significantly.

---

Deliverables for Fix 1.2:

• If model exists: Citation, preliminary results (1-2 pages), architecture diagram (1 page), INCITE allocation letter
• If model doesn't exist: New Specific Aim 1 (pre-training, 3-5 pages), revised budget (+$7-13M), fallback plan (1-2 pages), risk mitigation (1 page)

Estimated Effort: 1-2 weeks (if model exists) OR 3-4 weeks (if must add pre-training aim)

Impact on Score:

• Innovation dimension: 8.5 → 9.0 (+0.5 points) if model validated
• Approach dimension: 7.5 → 8.5 (+1.0 points) if feasibility de-risked
• Overall composite: 7.8 → 8.3-8.5 (+0.5 to +0.7 points)

---

1.3 ADD DETAILED 50-SITE COORDINATION PLAN [HIGH PRIORITY]

Current State: "50 sites" mentioned but no recruitment plan, no retention strategy, no governance structure

Required Fix:

Step 1: Site Recruitment Plan (3-5 pages)

Site Eligibility Criteria:

• Academic or clinical site with autism diagnostic capabilities
• MRI scanner (minimum 1.5T, pediatric imaging capabilities)
• ADOS-2 certified evaluator on staff (or willing to get certified)
• IRB capacity to approve multi-site study within 6-12 months
• Minimum patient volume: 60 DD diagnoses/year (to recruit 60 over 5 years)
• Genomics access: On-site lab OR partnership with external genomics core
• IT infrastructure: Secure data transfer (SFTP, federated learning server connection)

Recruitment Strategy:

• Phase 1 (Months 1-6): Recruit 10 "anchor sites" (well-established partners, high capacity)

  • Leverage PI's existing networks (ENIGMA, ABIDE contacts)
  • Target tier-1 academic medical centers (e.g., UCLA, Stanford, Yale, MGH in US; Oxford, KCL in UK; Seoul National University in Korea)
  • Deliverable: 10 site commitment letters

• Phase 2 (Months 7-12): Recruit 20 "core sites" (mix of academic + large community clinics)

  • Advertise at conferences (INSAR, IMFAR, OHBM, ACNP)
  • Direct outreach to autism research centers (SFARI grantees, Autism Centers of Excellence)
  • Deliverable: 20 site commitment letters

• Phase 3 (Months 13-18): Recruit 20 "diversity sites" (rural, low-resource, international)

  • Partner with global health organizations (WHO, UNICEF regional offices)
  • Offer resource-sharing (central genomics core, traveling EEG units)
  • Deliverable: 20 site commitment letters

Recruitment Incentives:

• Scientific: Co-authorship on main papers (ICMJE criteria), site-specific analyses for local publications
• Financial: $100K/site/5 years ($20K/year for staff time, patient recruitment)
• Training: Central IRB, ADOS-2 training, federated learning technical support
• Data: Sites retain access to their own data + aggregate de-identified global dataset (for secondary analyses)

Recruitment Timeline:

• Target: 50 sites recruited within 18 months
• Assumption: 20-30% attrition over 5 years → recruit 65 sites initially to ensure 50 complete
• Replacement strategy: Waitlist of 10-15 backup sites (if primary site drops, activate backup)

Step 2: Site Retention Plan (2-3 pages)

Retention Strategies:

• Regular communication: Monthly site PI calls, quarterly steering committee meetings
• Progress transparency: Real-time dashboards showing recruitment, data quality, federated model performance (by site)
• Recognition: Acknowledge top-performing sites (fastest recruitment, highest data quality) in newsletters
• Flexibility: Allow sites to pause recruitment if capacity issues (e.g., COVID-19-like disruptions)

Site Support:

• Central IRB: Single IRB protocol for all US sites (reduces local IRB burden from 12 months to 2-3 months)
• Data management training: 2-day on-site training (or virtual) for research coordinators
• Technical support: 24/7 federated learning server support, MRI quality control feedback
• Troubleshooting: Dedicated project manager for each region (US, Europe, Asia, Latin America, Africa)

Attrition Assumptions:

• Expected attrition: 20-30% over 5 years (typical for long studies)
• Reasons: PI leaves institution, loss of funding, IRB issues, loss of ADOS-2 staff
• Mitigation: Over-recruit to 65 sites, maintain waitlist of 10-15 backups

Step 3: Governance Structure (2-3 pages)

Steering Committee:

• Composition: PI (chair) + 5 site leads (1 per continent) + 2 advisory board members + NIH program officer (ex officio)
• Responsibilities:
  • Approve major protocol changes (e.g., add new modality, change inclusion criteria)
  • Review interim analyses (Bayesian adaptive design stopping rules)
  • Resolve site conflicts (e.g., data quality issues, authorship disputes)
• Meetings: Quarterly (in-person at conferences OR virtual)

Data Coordinating Center (DCC):

• Location: Host institution (PI's university)
• Staffing: 5 FTE (project manager, data manager, biostatistician, federated learning engineer, QC analyst)
• Responsibilities:
  • Data quality monitoring (MRI QC, genomics QC, outlier detection)
  • Federated learning server management (model aggregation, site-specific fine-tuning)
  • Statistical analyses (primary/secondary outcomes, interim analyses)
  • Regulatory compliance (IRB renewals, FDA reporting)

Publication Policy:

• Main Papers: ICMJE authorship (substantial contribution, draft/critical revision, final approval)
  • Anticipated: 10-15 main papers (all site PIs co-authors if recruited ≥10 participants)
• Site-Specific Papers: Sites can publish their own data (single-site analyses) with DCC co-authorship
• Authorship Order: Alphabetical by site OR contribution-based (decided by steering committee)

Step 4: Budget Breakdown (1-2 pages)

Current Budget: "5000백만원" ($50M) total - but no breakdown

Detailed Budget:

Category	Amount ($M)	Justification
Site Payments	$25M	50 sites × $100K/site × 5 years (patient recruitment, staff time, local IRB)
Data Coordinating Center	$5M	5 FTE × 5 years × $200K/FTE (salaries, benefits, overhead)
Compute (INCITE/Cloud)	$10M	Aurora pre-training ($5M) + DGX fine-tuning ($2M) + cloud backup ($3M)
Clinical Trial (pRCT)	$5M	10 sites × $500K/site (pRCT-specific costs, ADOS-2 assessments, regulatory)
Genomics	$2M	WES for 2,000 participants ($1,000/sample)
Regulatory/FDA	$2M	FDA submission ($500K), ISO 13485 QMS ($500K), regulatory consultant ($1M)
Contingency (20%)	$10M	Unexpected costs (site dropout, compute overruns, regulatory delays)
TOTAL	$59M	(Revised from $50M)

Justification for Budget Increase:

• Original $50M underestimated genomics ($2M) and contingency ($10M)
• Recommend: Request $60M OR de-scope to $50M (reduce sites to 30, genomics to n=1,000)

---

Deliverables for Fix 1.3:

• Site recruitment plan (3-5 pages): Eligibility, strategy, timeline, incentives
• Site retention plan (2-3 pages): Retention strategies, support, attrition assumptions
• Governance structure (2-3 pages): Steering committee, DCC, publication policy
• Budget breakdown (1-2 pages): Detailed line items totaling $50M (or revised to $60M)

Estimated Effort: 2-3 weeks

Impact on Score:

• Approach dimension: 7.5 → 8.2-8.5 (+0.7 to +1.0 points)
• Environment dimension: 7.8 → 8.5 (+0.7 points)
• Overall composite: 7.8 → 8.2-8.5 (+0.4 to +0.7 points)

---

TOTAL IMPACT OF PRIORITY 1 FIXES:

Before:

• Composite Score: 7.8/9.0 (Top 10-15%)
• Success Probability: 65-75%

After Priority 1:

• Composite Score: 8.3-8.7/9.0 (Top 5-8%)
• Success Probability: 80-90%

Estimated Total Effort: 6-10 weeks (parallel work on all 3 fixes)

Estimated Total Cost: $50K-100K (preliminary data collection, letter solicitation)

---

PRIORITY 2: HIGH-IMPACT IMPROVEMENTS (STRENGTHENS TOP 5%)

Definition: These improvements aren't mandatory, but significantly strengthen competitiveness and polish.

Estimated Impact: +0.3 to +0.5 points overall (8.3-8.7 → 8.6-9.0)

Timeline: 2-4 weeks additional effort

---

2.1 ADD STATISTICAL POWER CLUSTER ADJUSTMENT

Issue: Power calculations assume independent observations, but 50 sites = 50 clusters (non-independent)

Fix:

1. Estimate site-level ICC from ABIDE/ADHD-200:

   • Download ABIDE data (public, n=1,112, 17 sites)
   • Fit mixed-effects model: Diagnosis ~ Brain_Features + (1|Site)
   • Extract ICC = Var(Site) / [Var(Site) + Var(Residual)]
   • Expected ICC: 0.05-0.15 (typical for multi-site neuroimaging)

2. Calculate design effect:

   • Design effect = 1 + (m-1) × ICC
   • Where m = average cluster size = 60 participants/site
   • Example: ICC=0.10 → DE = 1 + 59×0.10 = 6.9

3. Recalculate effective sample size:

   • n_eff = n / DE = 3,000 / 6.9 = 435

4. Recalculate power:

   • Original: n=3,000, power >99% for d=0.5
   • Cluster-adjusted: n_eff=435, power = ? (use G*Power or simulation)
   • If power drops below 80%: Increase sample size OR reduce ICC (add covariates)

5. Add to proposal:

   • Section: "Cluster-Adjusted Power Analysis" (1-2 pages)
   • Table: Power for different ICC values (0.05, 0.10, 0.15)
   • Sensitivity: If power inadequate, mitigation strategies (increase n, add fixed effects for scanner type)

Deliverable: 1-2 pages added to "Statistical Methods" section

Effort: 1 week (download ABIDE, run analysis, add to proposal)

Impact: Approach score: 7.5 → 8.5 (+1.0 points)

---

2.2 ADD ALGORITHMIC FAIRNESS ANALYSIS PLAN

Issue: FDA requires performance stratified by demographic subgroups (21st Century Cures Act), but proposal lacks fairness plan

Fix:

Step 1: Define Demographic Subgroups

• Race/Ethnicity: White, Black/African American, Hispanic/Latino, Asian, Native American, Other/Mixed
• Sex: Male, Female
• Age: 0-2 years, 2-5 years, 5-10 years, 10-18 years
• Socioeconomic Status: Low (<$50K), Middle ($50K-$150K), High (>$150K)
• Geographic: Urban, Suburban, Rural

Step 2: Define Fairness Metrics

• Demographic Parity: P(Predicted ASD | Subgroup A) ≈ P(Predicted ASD | Subgroup B)
• Equal Opportunity: Sensitivity should be equal across subgroups (no group has lower true positive rate)
• Equalized Odds: Sensitivity AND specificity equal across subgroups
• Calibration: Predicted probabilities match observed outcomes across subgroups

Step 3: Fairness Analysis Plan

• Primary Fairness Metric: Equalized odds (FDA preference)
• Acceptable Disparity: ≤5 percentage points difference in sensitivity/specificity across subgroups
  • Example: Sensitivity in White = 95%, Black = 92% → 3-point gap (acceptable)
• If Disparity >5 Points: Apply bias mitigation
  • Re-weighting: Oversample underrepresented groups in training
  • Adversarial debiasing: Add fairness constraint to loss function
  • Group-specific thresholds: Optimize decision threshold per subgroup

Step 4: Add to Proposal

• Section: "Algorithmic Fairness and Health Equity" (2-3 pages)
• Table: Expected sample size per subgroup (to ensure adequate power for fairness analysis)
• Figure: Schematic of fairness analysis pipeline

Deliverable: 2-3 pages added to "Approach" section

Effort: 1 week

Impact: Approach score: +0.3, addresses FDA requirement

---

2.3 ADD MISSING MODALITY PERFORMANCE ANALYSIS

Issue: Real-world deployments will have missing modalities (e.g., no genomics due to cost). Performance degradation unclear.

Fix:

Step 1: Simulate Missing Modality Scenarios

• Scenario 1: All 5 modalities (sMRI, fMRI, EEG, genomics, digital) → Baseline AUC 0.92-0.95
• Scenario 2: 4 modalities (drop genomics, most expensive) → AUC = ?
• Scenario 3: 3 modalities (drop genomics + EEG) → AUC = ?
• Scenario 4: 2 modalities (imaging only: sMRI + fMRI) → AUC = ?
• Scenario 5: 1 modality (digital only, cheapest/most scalable) → AUC = ?

Step 2: Estimate Performance from Literature

• 5 modalities: 0.92-0.95 (proposed, multimodal synergy)
• 4 modalities: 0.90-0.92 (slight drop)
• 3 modalities: 0.88-0.90 (moderate drop)
• 2 modalities (imaging): 0.85-0.87 (CCTF benchmark: 0.82-0.87)
• 1 modality (digital): 0.88-0.90 (ADHD wearables: 0.89-0.95)

Step 3: Clinical Decision Thresholds

• Tier 1 screening (digital only): AUC 0.88-0.90 acceptable (high sensitivity, moderate specificity)
• Tier 2 confirmation (imaging + genomics): AUC 0.92-0.95 required (high sensitivity AND specificity)
• Minimum acceptable AUC: 0.85 (FDA Canvas Dx has 81.6% specificity, ~0.90 AUC estimated)

Step 4: Add to Proposal

• Section: "Missing Modality Robustness Analysis" (1-2 pages)
• Table: AUC across modality combinations (with 95% CI)
• Figure: Performance degradation curve (AUC vs. number of modalities)

Deliverable: 1-2 pages added to "Approach" section

Effort: 1 week

Impact: Approach score: +0.2, demonstrates real-world robustness

---

2.4 REVISE IMPACT PROJECTIONS (CONSERVATIVE ESTIMATES)

Issue: Some impact metrics are inflated and hurt credibility

Fix:

Inflated Claim #1: "40-60 Nature/Science papers"

• Reality Check: Large consortia publish 5-10 Nature/Science papers over 10 years (ENIGMA: ~10, HCP: ~8)
• Revised Estimate: 10-15 high-impact papers (Nature, Science, Nature Medicine, JAMA, Lancet)
  • 5 main outcomes papers (diagnosis, subtyping, genomics, causal inference, pRCT)
  • 5 methods papers (foundation model, federated learning, multimodal fusion)
  • 5 secondary analyses (sex differences, developmental trajectories, treatment response)
• Plus: 30-40 total papers (including mid-tier journals: NeuroImage, Biological Psychiatry, Autism Research)

Inflated Claim #2: "10-20% global market share within 5 years of FDA approval"

• Reality Check: Canvas Dx (4 years post-FDA): estimated 5-10% US market penetration
• Revised Estimate: 5-10% global market share within 5 years
  • Optimistic: 10% ($50-80M annual revenue)
  • Realistic: 5% ($25-40M annual revenue)
  • Pessimistic: 2-3% ($10-24M annual revenue, niche player)

Inflated Claim #3: "조기중재로 발달 전환 2-3배 개선"

• Basis: Early intervention literature shows 20-50% symptom reduction (not 2-3× improvement)
• Revised Estimate: "30-50% improvement in developmental outcomes (IQ, adaptive functioning, symptom severity) compared to standard care"

Deliverable: Revise "Expected Outcomes" section (1-2 pages) with conservative estimates

Effort: 1-2 days

Impact: Impact score: 8.2 → 8.5 (+0.3 points, improved credibility)

---

2.5 ADD ETHICS AND REGULATORY PLANNING

Issue: Ethical issues (early diagnosis, genetic counseling) and FDA requirements (risk management) are under-addressed

Fix:

Step 1: Ethical Framework (2-3 pages)

Issue #1: Early Diagnosis at 6-12 Months - Ethical Considerations

• Pro: Enables early intervention during peak neuroplasticity
• Con: Risk of labeling, family anxiety, false positives
• Mitigation:
  • Genetic counseling for all families (pre-test and post-test)
  • Clear communication: "High risk" ≠ "definite diagnosis" (only ADOS-2 at 24 months is diagnostic)
  • Psychosocial support: Parent support groups, mental health resources

Issue #2: Incidental Findings (Genomics + Imaging)

• ACMG SF v3.0: Must report actionable secondary findings (cancer genes, cardiac genes)
• Plan:
  • Pre-test counseling: Families opt-in to receive incidental findings
  • Reporting protocol: Clinical geneticist reviews all WES, reports ACMG SF variants
  • Follow-up: Refer to appropriate specialist (oncology, cardiology)

Issue #3: Algorithmic Bias and Health Equity

• Risk: AI may underperform in minority populations (as addressed in Fix 2.2)
• Mitigation: Fairness analysis, bias mitigation, diverse recruitment
• Equity Plan: Low-resource sites participate via resource-sharing (central genomics core)

Step 2: FDA Risk Management (ISO 14971) (2-3 pages)

Hazard Analysis:

Hazard	Severity	Likelihood	Risk Level	Mitigation
False Positive	Moderate (family anxiety, unnecessary intervention)	Medium (10-15% expected)	Medium	Pre-test counseling, confirmatory ADOS-2 at 24 months
False Negative	High (missed diagnosis, delayed treatment)	Low (5-10% expected)	Medium-High	Longitudinal monitoring (catch late-onset ASD), secondary screening at 18-24 months
Model Bias	High (underperforms in minorities → health inequity)	Medium (if not mitigated)	High	Fairness analysis (Fix 2.2), diverse recruitment, bias mitigation algorithms
Data Breach	High (genomic data leak → re-identification)	Low (strong encryption, federated learning)	Medium	Differential privacy, homomorphic encryption, HIPAA compliance, cybersecurity audits
Algorithm Drift	Moderate (performance degrades over time as population shifts)	Medium (without monitoring)	Medium	Post-market surveillance (continuous performance monitoring), model updates

Risk Mitigation Strategies:

• Design Controls: Fairness constraints in model training, differential privacy (ε=1.0)
• Verification/Validation: pRCT (n=500, 10 sites), external validation (50 sites)
• Post-Market Surveillance: Real-world performance monitoring (quarterly reports to FDA)

Step 3: Post-Market Surveillance Plan (1-2 pages)

FDA Requirement: Continuous monitoring of AI/ML device performance post-deployment

Plan:

• Data Collection: All clinical deployments report outcomes (predicted diagnosis vs. ADOS-2 gold standard)
• Frequency: Quarterly performance reports (sensitivity, specificity, AUC) to FDA
• Thresholds: If sensitivity <90% or specificity <85% for 2 consecutive quarters → trigger investigation
• Model Updates: Algorithm Change Protocol (pre-specified by FDA) allows updates without new clearance (if performance stays within bounds)

Deliverables:

• Ethical framework (2-3 pages)
• ISO 14971 risk management file (2-3 pages)
• Post-market surveillance plan (1-2 pages)

Effort: 2 weeks

Impact: Approach score: +0.3, addresses FDA requirements proactively

---

TOTAL IMPACT OF PRIORITY 2 FIXES:

Before Priority 2:

• Composite Score: 8.3-8.7/9.0 (after Priority 1)

After Priority 2:

• Composite Score: 8.6-9.0/9.0 (Top 3-5%, excellent positioning)
• Success Probability: 85-95%+

Estimated Total Effort (Priority 2): 4-6 weeks

---

PRIORITY 3: POLISH FOR TOP 1-3% (OPTIONAL)

Definition: These are refinements that elevate a strong proposal to exceptional. Not required for funding, but increase probability of "Outstanding" rating.

Estimated Impact: +0.1 to +0.3 points overall (8.6-9.0 → 8.7-9.0+, capped)

Timeline: 2-3 weeks additional effort

---

3.1 ADD MODEL INTERPRETABILITY VALIDATION STUDY

Proposal: Conduct clinician validation of AI explanations (n=10-15 child psychiatrists)

Method:

• Present 20 sample cases (10 ASD, 10 TD) with AI predictions + explanations (attention maps, SHAP values)
• Ask clinicians: "Do you understand the AI's reasoning?" (5-point Likert scale)
• Ask: "Do you agree with the AI's prediction?" (Yes/No/Uncertain)
• Target: ≥70% clinicians rate explanations as "understandable" (4-5 on Likert), ≥80% agree with predictions

Deliverable: 1-2 pages added to "Interpretability" section, 1 figure showing clinician comprehension results

Effort: 2-3 weeks (recruit clinicians, prepare materials, run study, analyze)

Impact: Approach score: +0.2, demonstrates clinical usability

---

3.2 ADD DIFFERENTIAL PRIVACY SENSITIVITY ANALYSIS

Proposal: Show performance vs. privacy trade-off (ε=0.1, 1.0, 10, ∞)

Method:

• Simulate federated learning on ABIDE with different DP noise levels
• Measure: AUC vs. ε (privacy budget)
• Expected: ε=∞ (no DP): AUC 0.90, ε=10: AUC 0.89, ε=1.0: AUC 0.87, ε=0.1: AUC 0.82
• Recommendation: ε=1.0 balances privacy (strict) and performance (acceptable)

Deliverable: 1 page added to "Federated Learning" section, 1 figure showing AUC vs. ε

Effort: 1 week

Impact: Approach score: +0.1, demonstrates rigor

---

3.3 ADD CPT CODE REIMBURSEMENT STRATEGY

Proposal: Plan for insurance reimbursement (CPT code application)

Method:

• Identify existing CPT codes that could apply (e.g., 96127 brief emotional/behavioral assessment)
• Propose new CPT code: "AI-assisted autism diagnostic evaluation" (estimated reimbursement $400-600)
• Timeline: Apply to AMA CPT Editorial Panel (Year 6-7, after pRCT results)
• Payer engagement: Pilot contracts with 2-3 large insurers (Blue Cross, Aetna, UnitedHealthcare)

Deliverable: 1-2 pages added to "Commercialization" section

Effort: 1 week

Impact: Impact score: +0.1, demonstrates real-world sustainability

---

3.4 ADD QUALITY CONTROL PLAN FOR MULTI-SITE IMAGING

Proposal: Detailed MRI/EEG quality control protocol (following ENIGMA best practices)

Method:

• MRI QC:
  • Automated: FreeSurfer QC scripts (Euler number, surface holes, contrast-to-noise ratio)
  • Manual: Visual inspection of 10% random sample (trained raters, inter-rater reliability κ≥0.80)
  • Exclusion criteria: Motion >3mm translation, severe artifacts, failed segmentation
• EEG QC:
  • Automated: Artifact detection (eye blinks, muscle artifacts), signal-to-noise ratio
  • Manual: Visual inspection of event-related potentials (N170, ERN)
  • Exclusion criteria: >30% trials rejected, low SNR (<3 dB)

Deliverable: 2-3 pages added to "Data Quality Assurance" section

Effort: 1 week

Impact: Approach score: +0.1, demonstrates rigor

---

TOTAL IMPACT OF PRIORITY 3 FIXES:

Before Priority 3:

• Composite Score: 8.6-9.0/9.0 (after Priority 1 + 2)

After Priority 3:

• Composite Score: 8.7-9.0/9.0 (Top 1-3%, exceptional)
• Success Probability: 90-95%+

Estimated Total Effort (Priority 3): 3-4 weeks

---

IMPLEMENTATION TIMELINE: 8-WEEK CRITICAL PATH

Week 1-2: Priority 1A (Investigators)

• Identify and recruit PI + Co-Is (6 investigators)
• Draft biosketches (5 pages each)
• Output: Named investigators + CVs

Week 3-4: Priority 1B (Preliminary Data)

• Collect preliminary data (pilot multimodal fusion OR FL simulation OR LoRA fine-tuning)
• Output: 5-10 page preliminary data report

Week 5-6: Priority 1C (INCITE Model)

• Clarify NeuroX-Fusion 130B status (cite + validate OR add pre-training aim)
• Secure INCITE allocation letter (or confirm model access)
• Output: Model validation OR new Specific Aim 1

Week 7: Priority 1D (Site Coordination)

• Write site recruitment/retention/governance plans
• Develop detailed budget breakdown
• Output: 8-10 pages of operational details

Week 8: Priority 2 (High-Impact Fixes)

• Add cluster power analysis
• Add fairness analysis plan
• Add missing modality analysis
• Revise impact projections
• Add ethics + risk management
• Output: 10-15 pages of additional content

Week 9-10 (Optional): Priority 3 (Polish)

• Interpretability validation
• DP sensitivity
• CPT code plan
• QC plan
• Output: 5-8 pages of refinements

Week 11-12 (Optional): Final Assembly

• Integrate all fixes into main proposal
• Proofread, format, check page limits
• Solicit external feedback (advisory board, mock review)
• Output: Final polished proposal

---

FINAL RECOMMENDATIONS SUMMARY

Must-Do (Mandatory for Top 5%)

1. Name investigators with track records (PI + 5 Co-Is, CVs, preliminary data, letters of support)
2. Clarify INCITE model status (cite + validate OR add pre-training aim + fallback)
3. Add 50-site coordination plan (recruitment, retention, governance, budget breakdown)

Timeline: 6-8 weeks Impact: 7.8 → 8.3-8.7 (Top 5-8%)

Highly Recommended (Strengthens Top 5%)

4. Cluster-adjusted power analysis
5. Algorithmic fairness plan
6. Missing modality analysis
7. Revise impact projections (conservative)
8. Ethics + risk management + post-market surveillance

Timeline: +4 weeks (total 10-12 weeks) Impact: 8.3-8.7 → 8.6-9.0 (Top 3-5%)

Optional Polish (Top 1-3%)

9. Interpretability validation
10. DP sensitivity analysis
11. CPT code reimbursement plan
12. MRI/EEG QC plan

Timeline: +3 weeks (total 13-15 weeks) Impact: 8.6-9.0 → 8.7-9.0+ (Top 1-3%, exceptional)

---

CONCLUSION: YOUR REVOLUTIONARY PROPOSAL IS WITHIN REACH

Current State:

• This proposal has outstanding scientific foundations (innovation, rigor, impact)
• It is currently in top 10-15% tier (fundable, but not top 5%)
• 3 critical barriers prevent top 5% success: Investigators, INCITE model, site coordination

With Focused Effort:

• 6-8 weeks of dedicated work on Priority 1 → Top 5-8% tier
• 10-12 weeks total with Priority 1+2 → Top 3-5% tier
• 13-15 weeks total with all priorities → Top 1-3% tier (exceptional)

This proposal can become FIELD-DEFINING with the right team, preliminary data, and operational details.

The science is revolutionary. The execution plan needs strengthening. You have a clear roadmap to top 5% success.

Recommended Next Steps:

1. Week 1: Identify PI candidate (reach out to ENIGMA/ABIDE leaders)
2. Week 2: Recruit Co-Investigators (send invitations with 1-page project summary)
3. Week 3-4: Collect preliminary data (even n=50 pilot shows commitment)
4. Week 5-8: Implement all Priority 1 fixes
5. Week 9-12: Implement Priority 2 fixes
6. Week 13: Submit exceptional proposal

You have the potential for a revolutionary, field-defining grant. Execute this roadmap, and success is highly likely.