from pptx import Presentation
from pptx.util import Inches, Pt
from pptx.enum.shapes import MSO_SHAPE, MSO_CONNECTOR
from pptx.enum.text import PP_ALIGN, MSO_ANCHOR
from pptx.dml.color import RGBColor
from pptx.enum.dml import MSO_THEME_COLOR
from pptx.enum.shapes import MSO_AUTO_SHAPE_TYPE
from pptx.enum.text import MSO_AUTO_SIZE
from reportlab.pdfgen import canvas
from reportlab.lib.pagesizes import landscape
from reportlab.lib.colors import HexColor, white
from reportlab.pdfbase.pdfmetrics import stringWidth
import json, base64, os, textwrap

OUT='/tmp/workspace/ai-medicine-presentation/output'
AS='/tmp/workspace/ai-medicine-presentation/assets'
os.makedirs(OUT, exist_ok=True)
W,H=13.333,7.5
NAVY='0B1F33'; TEAL='00A6A6'; CYAN='36C5F0'; BLUE='2056A5'; SKY='DFF7FA'; PALE='F4F8FB'; INK='162B3A'; GREY='5D7180'; RED='D65353'; GOLD='F6B34B'; WHITE='FFFFFF'; LINE='C9D7E0'
def rgb(x): return RGBColor.from_string(x)

def rect(s,x,y,w,h,fill, radius=False, line=None):
    sh=s.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE if radius else MSO_SHAPE.RECTANGLE, Inches(x), Inches(y), Inches(w), Inches(h))
    sh.fill.solid(); sh.fill.fore_color.rgb=rgb(fill)
    sh.line.color.rgb=rgb(line or fill)
    return sh

def txt(s,x,y,w,h,text,size=20,color=INK,bold=False,align=PP_ALIGN.LEFT, font='Aptos Display', margin=.06):
    tb=s.shapes.add_textbox(Inches(x), Inches(y), Inches(w), Inches(h)); tf=tb.text_frame; tf.clear(); tf.word_wrap=True
    tf.margin_left=tf.margin_right=Inches(margin); tf.margin_top=tf.margin_bottom=Inches(margin)
    p=tf.paragraphs[0]; p.alignment=align; p.space_after=Pt(0); p.space_before=Pt(0)
    for i,line in enumerate(str(text).split('\n')):
        if i: p=tf.add_paragraph(); p.alignment=align; p.space_after=Pt(0)
        p.text=line
        for r in p.runs:
            r.font.name=font; r.font.size=Pt(size); r.font.bold=bold; r.font.color.rgb=rgb(color)
    tf.vertical_anchor=MSO_ANCHOR.MIDDLE
    return tb

def bullet(s,x,y,w,items,size=17,accent=TEAL, gap=.48):
    for i,t in enumerate(items):
        cy=y+i*gap
        rect(s,x,cy+.15,.12,.12,accent,True)
        txt(s,x+.24,cy,w-.24,.4,t,size,INK)

def title(s,n,kicker,title,sub=''):
    rect(s,0,0,W,.16,TEAL); txt(s,.55,.35,1.2,.3,kicker.upper(),10,TEAL,True); txt(s,.55,.65,11.9,.6,title,27,NAVY,True)
    if sub: txt(s,.57,1.29,11.6,.35,sub,11,GREY)
    txt(s,12.25,7.05,.5,.22,f'{n:02}',10,TEAL,True,PP_ALIGN.RIGHT)

def footer(s,ref=''):
    rect(s,.55,6.86,11.9,.012,LINE)
    if ref: txt(s,.55,6.92,11.3,.22,ref,7,GREY)

def card(s,x,y,w,h,heading,body,accent=TEAL):
    rect(s,x,y,w,h,WHITE,True,LINE); rect(s,x,y,.09,h,accent,False,accent)
    txt(s,x+.22,y+.16,w-.38,.35,heading,15,NAVY,True); txt(s,x+.22,y+.58,w-.38,h-.68,body,11.5,INK)

def arrow(s,x1,y1,x2,y2,color=TEAL):
    ln=s.shapes.add_connector(MSO_CONNECTOR.STRAIGHT,Inches(x1),Inches(y1),Inches(x2),Inches(y2)); ln.line.color.rgb=rgb(color); ln.line.width=Pt(2); ln.line.end_arrowhead=True

def notes(slide, text):
    tf=slide.notes_slide.notes_text_frame
    tf.text=text

def imagefile():
    data=json.load(open(AS+'/images.json'))
    paths=[]
    for i,it in enumerate(data):
        b64=it.get('base64','').split(',',1)[-1]
        p=f'{AS}/image{i}.jpg'; open(p,'wb').write(base64.b64decode(b64)); paths.append(p)
    return paths

prs=Presentation(); prs.slide_width=Inches(W); prs.slide_height=Inches(H); blank=prs.slide_layouts[6]
imgs=imagefile()
def new(bg=PALE):
    s=prs.slides.add_slide(blank); rect(s,0,0,W,H,bg); return s

# 1
s=new(NAVY); rect(s,0,0,.18,H,TEAL); rect(s,7.75,0,5.58,H,'103B5B');
txt(s,.72,.78,2.5,.3,'RESEARCH METHODOLOGY ELECTIVE',12,CYAN,True)
txt(s,.7,1.38,6.45,1.5,'Artificial\nIntelligence in Medicine',34,WHITE,True)
txt(s,.75,3.2,5.8,.55,'From algorithms to accountable clinical and research use',17,'B6D5E6')
# neural/network motif
for x,y in [(9,1.4),(10.5,1.0),(11.7,1.9),(9.5,3),(11,3.6),(8.55,4.25),(10.1,5.0),(12,5.1)]:
    rect(s,x,y,.36,.36,CYAN,True)
for a,b in [((9.2,1.58),(10.68,1.18)),((10.68,1.18),(11.88,2.08)),((9.2,1.58),(9.68,3.18)),((9.68,3.18),(11.18,3.78)),((8.73,4.43),(10.28,5.18)),((11.18,3.78),(12.18,5.28))]: arrow(s,*a,*b,'55B7D1')
txt(s,.75,6.15,5.5,.5,'Mayur Lokhande  |  Department of Community Medicine / PSM',13,WHITE)
notes(s,'Open with the central premise: AI can support clinical work and research, but it does not remove professional responsibility. Faculty question: Is AI already used in hospitals? Yes, particularly in imaging, triage, monitoring, workflow and approved software devices.')
#2
s=new(); title(s,2,'Roadmap','Learning objectives','By the end, you should be able to explain AI use and critically appraise an AI medical study.');
for i,(h,b) in enumerate([('Explain','AI, ML, deep learning and generative AI'),('Apply','clinical and public-health applications'),('Evaluate','datasets, validation and performance metrics'),('Govern','bias, ethics, privacy and accountable use')]): card(s,.75+(i%2)*6.05,2.0+(i//2)*1.9,5.55,1.45,h,b,[TEAL,BLUE,GOLD,RED][i])
footer(s,'Scope: MBBS level, clinical relevance, research-methodology focus')
notes(s,'Signpost the flow: concepts, clinical value, research methods, then critical appraisal and ethics. Invite the audience to judge AI by clinical usefulness, not novelty.')
#3
s=new(); title(s,3,'Foundations','What is artificial intelligence?','AI performs tasks associated with aspects of human intelligence by processing data through algorithms.');
for i,(h,b) in enumerate([('Data','Images, ECGs, notes, labs, genomics'),('Algorithm','Rules or learned mathematical patterns'),('Output','Classification, prediction, prioritisation or generated content')]):
    x=.8+i*4.1; rect(s,x,2.25,3.4,1.25,['DFF7FA','E8F0FC','FDF3DD'][i],True); txt(s,x+.18,2.48,3,.3,h,18,NAVY,True,PP_ALIGN.CENTER); txt(s,x+.2,2.92,3, .35,b,11,INK,False,PP_ALIGN.CENTER)
    if i<2: arrow(s,x+3.45,2.86,x+4.0,2.86)
txt(s,.85,4.55,11.5,.55,'In medicine, a model should generate an interpretable, context-aware input to a clinician - not an autonomous medical decision.',20,NAVY,True,PP_ALIGN.CENTER)
footer(s,'Harrison’s Principles of Internal Medicine, 22e (2025), Ch. 501')
notes(s,'Define AI broadly. A model can identify patterns, but it does not possess clinical accountability or understand the patient’s values. Interesting fact: many deployed medical AI systems are narrow, task-specific systems.')
#4
s=new(); title(s,4,'Foundations','AI → machine learning → deep learning','A hierarchy, not interchangeable terms. Generative AI is defined by what it produces, not by a separate level of the hierarchy.');
levels=[('Artificial intelligence','Broad field: systems that perform intelligent tasks','0B1F33'),('Machine learning','Algorithms learn statistical patterns from examples','2056A5'),('Deep learning','ML using multi-layer neural networks','00A6A6'),('Generative AI / LLMs','Generate text, images, code or summaries','F6B34B')]
for i,(h,b,c) in enumerate(levels):
    x=.85+i*3.05; w=2.75; rect(s,x,2.2,w,2.35,c,True); txt(s,x+.15,2.5,w-.3,.45,h,17,WHITE if i<3 else NAVY,True,PP_ALIGN.CENTER); txt(s,x+.18,3.25,w-.36,.75,b,11,WHITE if i<3 else NAVY,False,PP_ALIGN.CENTER)
    if i<3: arrow(s,x+w,3.35,x+3.0,3.35,'90C8DB')
footer(s,'Generative AI requires output verification: it may produce fluent but incorrect content.')
notes(s,'Emphasise that ML learns from examples; deep learning is a subset that can learn complex representations, especially from images and signals. LLMs can assist with language tasks but may hallucinate facts or references.')
#5
s=new(); title(s,5,'Clinical need','Why AI in medicine?','The opportunity arises from high-volume, heterogeneous data and time-sensitive clinical decisions.');
items=[('Detect','subtle patterns in imaging, pathology, ECG and retinal photographs'),('Predict','risk of deterioration, readmission, disease progression or treatment response'),('Prioritise','worklists and early-warning alerts when validated in the target setting'),('Personalise','integrate clinical, molecular and longitudinal data'),('Learn','support research, surveillance and medical education')]
for i,(a,b) in enumerate(items): card(s,.85+(i%3)*4.1,2.0+(i//3)*1.8,3.72,1.35,a,b,[TEAL,BLUE,GOLD,RED,TEAL][i])
footer(s,'Potential does not equal outcome benefit: impact must be tested prospectively.')
notes(s,'Frame the unmet need: clinicians face more data than can be manually processed. The question is not whether an algorithm is accurate in isolation, but whether it improves care without new harms.')
#6
s=new(); title(s,6,'Clinical applications','Where does AI fit in medicine?','Examples should be understood as decision support, with performance depending on data, task and clinical setting.');
apps=[('Radiology','X-ray, CT and MRI triage / detection'),('Pathology','digital-slide screening and quantification'),('Ophthalmology','retinal image screening'),('Cardiology','ECG interpretation and risk prediction'),('Therapeutics','treatment planning, robotics, response prediction'),('Drug development','target identification, screening and trial operations'),('Monitoring','wearables, remote signals and deterioration alerts'),('Public health','surveillance, forecasting and resource allocation')]
for i,(h,b) in enumerate(apps): card(s,.65+(i%4)*3.16,1.85+(i//4)*1.6,2.9,1.22,h,b,[TEAL,BLUE,GOLD,RED][i%4])
footer(s,'FDA maintains a public list of AI-enabled medical devices; regulatory clearance is task-specific, not a blanket endorsement.')
notes(s,'Use one application only as an example at a time. Faculty question: Can AI replace radiologists? No. It may support defined image-analysis tasks, but radiology involves clinical context, communication, quality assurance and responsibility.')
#7
s=new(); title(s,7,'Research methodology','AI across the research pathway','AI can reduce repetitive work, but each step needs human methodological oversight.');
steps=['Question','Literature review','Design','Data capture','Analysis','Interpretation','Writing & reporting']
for i,h in enumerate(steps):
    x=.5+i*1.82; rect(s,x,2.35,1.52,1.05,[TEAL,BLUE,'4789B8',GOLD,'D65353',TEAL,BLUE][i],True); txt(s,x+.06,2.65,1.4,.36,h,12,WHITE,True,PP_ALIGN.CENTER)
    if i<6: arrow(s,x+1.54,2.88,x+1.77,2.88,'8AAFC1')
for x,h,b in [(1.1,'Useful tasks','screening abstracts, NLP extraction, clustering, prediction, visualisation'),(7.0,'Non-delegable tasks','question validity, consent, protocol decisions, causal inference, interpretation, authorship')]: card(s,x,4.35,5.25,1.35,h,b,TEAL if x<5 else RED)
footer(s,'AI output is a draft or decision-support input, not a substitute for a protocol, statistician or investigator judgment.')
notes(s,'Make the methodology link explicit. AI can help generate ideas, screen literature and process data, but cannot verify truth or replace governance. Any AI-generated question needs a human check for relevance, feasibility and ethics.')
#8
s=new(); title(s,8,'Data','What data does a medical AI study use?','Data provenance and labels are as important as the model.');
card(s,.8,1.95,3.7,2.1,'Structured','Age, sex, vitals, diagnoses, medications, laboratory values, coded outcomes',TEAL)
card(s,4.82,1.95,3.7,2.1,'Unstructured','Clinical notes, images, histopathology, ECG/waveforms, audio, video',BLUE)
card(s,8.84,1.95,3.7,2.1,'Multi-modal','Linking text + images + signals + longitudinal records may add context, but increases governance demands',GOLD)
rect(s,1.2,4.85,10.9,.85,'E9F7F7',True); txt(s,1.45,5.06,10.4,.35,'Data quality checklist: representative population  •  clear time window  •  reliable reference standard  •  complete metadata  •  privacy protection',14,NAVY,True,PP_ALIGN.CENTER)
footer(s,'Dataset shift occurs when the deployment population, workflow or equipment differs from development data.')
notes(s,'Explain that model performance can fall when scanners, coding practices, prevalence or patient demographics change. Labels should be derived from credible reference standards, not convenience labels alone.')
#9
s=new(); title(s,9,'Study design','Research methodology for an AI study','A transparent protocol is needed before model development.');
flow=[('1','Problem & users'),('2','Question & outcome'),('3','Design & population'),('4','Data & reference standard'),('5','Model development'),('6','Internal + external validation'),('7','Clinical impact & reporting')]
for i,(n,h) in enumerate(flow):
    x=.7+i*1.8; rect(s,x,2.2,1.42,1.2,TEAL if i in [0,6] else BLUE,True); txt(s,x+.1,2.38,1.22,.24,n,11,CYAN,True,PP_ALIGN.CENTER); txt(s,x+.1,2.68,1.22,.45,h,11,WHITE,True,PP_ALIGN.CENTER)
    if i<6: arrow(s,x+1.43,2.8,x+1.76,2.8,'A1C8D6')
card(s,1.1,4.45,5.15,1.18,'Reporting','Pre-specify outcomes and analysis; use transparent reporting frameworks such as TRIPOD-AI / CONSORT-AI where applicable.',GOLD)
card(s,7.0,4.45,5.15,1.18,'Clinical question','Does the system improve a patient-relevant process or outcome compared with usual care, not only its AUROC?',RED)
footer(s,'Frameworks evolve. Check the current EQUATOR Network guidance for the study type.')
notes(s,'Point out that clinical evaluation comes after technical development. A high AUROC is not sufficient evidence of benefit. Methods should state intended use, user, setting, comparator and outcome.')
#10
s=new(); title(s,10,'Study design','Choosing the right evaluation design','The evaluation design must match the intended clinical claim.');
for i,(h,b,c) in enumerate([('Retrospective diagnostic study','Compare model output with a reference standard; high risk of dataset bias.',TEAL),('Prospective cohort / silent trial','Run model in real workflow without influencing care; assess calibration, drift and safety.',BLUE),('Randomised impact study','Compare AI-assisted care with usual care for workflow or patient outcomes.',GOLD),('Systematic review','Synthesise heterogeneous studies cautiously; assess bias, validation and applicability.',RED)]): card(s,.8+(i%2)*6.0,1.9+(i//2)*2.0,5.45,1.5,h,b,c)
footer(s,'Example: AI-assisted chest X-ray triage should be assessed against a defined reference standard and workflow endpoint.')
notes(s,'Diagnostic accuracy answers a different question from patient benefit. For an alert system, a prospective evaluation is important because clinician response and alert burden affect real-world usefulness.')
#11
s=new(); title(s,11,'Question formulation','PICO + FINER: an AI research question','AI can help draft questions, but the investigator must verify feasibility and scientific value.');
for i,(h,b,c) in enumerate([('P','Patients undergoing chest X-ray',TEAL),('I','AI-assisted interpretation',BLUE),('C','Usual / clinician interpretation',GOLD),('O','Diagnostic accuracy and time-to-prioritisation',RED)]): card(s,.9+i*3.08,1.9,2.72,1.4,h,b,c)
rect(s,.9,3.8,11.5,1.1,'E9F7F7',True); txt(s,1.2,4.06,10.9,.55,'Does AI-assisted chest-X-ray interpretation improve diagnostic performance or worklist prioritisation versus usual interpretation in the intended clinical setting?',16,NAVY,True,PP_ALIGN.CENTER)
txt(s,1.05,5.45,11,.3,'FINER check: Feasible • Interesting • Novel • Ethical • Relevant',18,TEAL,True,PP_ALIGN.CENTER)
footer(s,'Do not assume “AI” is the intervention: specify the user, workflow, comparator and outcome.')
notes(s,'Give the PICO example, then use FINER. AI may help brainstorm synonyms or map concepts, but it cannot establish novelty without verified literature searching and cannot resolve ethical feasibility.')
#12
s=new(); title(s,12,'Sampling','Population, sampling and sample size','Representativeness and outcome prevalence affect precision, fairness and generalisability.');
for i,(h,b) in enumerate([('Define target population','Inclusion/exclusion criteria, care setting, period and intended users.'),('Plan sampling','Avoid convenience-only sampling; preserve temporal and site diversity.'),('Estimate sample size','For diagnostic models, consider sensitivity/specificity precision, prevalence and events.'),('Audit subgroups','Pre-specify performance checks by relevant demographic and clinical groups.')]): card(s,.85+(i%2)*6.0,1.85+(i//2)*1.85,5.45,1.4,h,b,[TEAL,BLUE,GOLD,RED][i])
rect(s,.9,5.7,11.5,.45,'FFF4E0',True); txt(s,1.15,5.78,11,.22,'Low-prevalence outcomes can yield deceptively high accuracy. Report sensitivity, specificity, PPV and NPV in context.',12,NAVY,True,PP_ALIGN.CENTER)
footer(s,'PPV and NPV depend on prevalence; accuracy alone can mislead.')
notes(s,'A model trained in a tertiary hospital may not generalise to primary care. Accuracy can look excellent when most patients do not have disease. Always ask who is missing from the data.')
#13
s=new(); title(s,13,'Validation','Training, validation and test data','Keep final evaluation data independent from all model-development decisions.');
for i,(h,b,c) in enumerate([('Training set','Learns patterns from data',TEAL),('Validation set','Tunes model / chooses settings',BLUE),('Test set','Final locked performance estimate',GOLD),('External test','Different site, time or population',RED)]):
    x=.8+i*3.1; rect(s,x,2.05,2.7,1.45,c,True); txt(s,x+.15,2.3,2.4,.32,h,17,WHITE if i<3 else WHITE,True,PP_ALIGN.CENTER); txt(s,x+.15,2.77,2.4,.3,b,11,WHITE,False,PP_ALIGN.CENTER)
    if i<3: arrow(s,x+2.72,2.77,x+3.02,2.77,'91C2D5')
card(s,1.1,4.65,5.2,1.15,'Avoid leakage','Never allow the same patient, image series or future information to contaminate test data.',RED)
card(s,7.0,4.65,5.2,1.15,'Why external validation?','It tests transportability across clinical settings and is more informative than internal performance alone.',TEAL)
footer(s,'Temporal and geographic external validation are especially useful before deployment.')
notes(s,'Define data leakage with a simple example: slices from the same CT patient must not land in train and test sets. External validation asks whether the result is likely to travel.')
#14
s=new(); title(s,14,'Performance','Measuring AI performance','Use a set of metrics tied to the clinical task, disease prevalence and consequences of error.');
metrics=[('Sensitivity','Finds disease among people with disease'),('Specificity','Correctly excludes disease among people without disease'),('PPV / NPV','Probability an output is correct in this prevalence'),('AUROC','Discrimination across thresholds'),('Calibration','Agreement between predicted and observed risk'),('Clinical utility','Net benefit, workflow and patient outcomes')]
for i,(h,b) in enumerate(metrics): card(s,.72+(i%3)*4.18,1.78+(i//3)*1.7,3.8,1.25,h,b,[TEAL,BLUE,GOLD,RED,TEAL,BLUE][i])
footer(s,'Technical accuracy ≠ clinical usefulness ≠ improved patient outcomes.')
notes(s,'Use screening as an example: sensitivity may be prioritised to miss fewer cases, but false positives can overburden services. Calibration matters when the number itself guides risk-based decisions.')
#15
s=new(); title(s,15,'Diagnostic accuracy','Confusion matrix: read it clinically','A test result is meaningful only in relation to a reference standard and the population being tested.');
# table
x,y=.95,1.85; cw=[2.7,2.85,2.85]; rh=.68
labels=[['','Disease +','Disease −'],['AI positive','True positive','False positive'],['AI negative','False negative','True negative']]
for r in range(3):
 for c in range(3):
  col=NAVY if r==0 or c==0 else ['DFF7FA','FFF0EF','FFF0EF','E8F8F4'][ (r-1)*2+(c-1)]
  rect(s,x+sum(cw[:c]),y+r*rh,cw[c],rh,col,False,WHITE); txt(s,x+sum(cw[:c])+.05,y+r*rh+.12,cw[c]-.1,.35,labels[r][c],14,WHITE if r==0 or c==0 else INK,True,PP_ALIGN.CENTER)
card(s,9.2,1.85,3.0,2.04,'Simple example','Among 100 patients with 10 disease cases, 90% sensitivity detects 9 cases; 95% specificity creates ~5 false positives among 90 non-cases.',GOLD)
rect(s,.95,4.55,11.25,.9,'E9F7F7',True); txt(s,1.15,4.77,10.85,.45,'Sensitivity = TP / (TP + FN)     •     Specificity = TN / (TN + FP)     •     PPV & NPV change with prevalence',14,NAVY,True,PP_ALIGN.CENTER)
footer(s,'Reference standard quality defines whether “true” and “false” are meaningful.')
notes(s,'Walk through why low prevalence reduces PPV even when sensitivity and specificity are high. Faculty question: Why is sensitivity important in screening? It helps reduce missed disease, but the cost of false positives still matters.')
#16
s=new(); title(s,16,'Fairness','Bias in AI: a medical example','A model can reproduce inequity if its data or labels are incomplete, unrepresentative or structurally biased.');
steps=[('Training data','Under-represents a population'),('Model','Learns patterns mainly from represented groups'),('Deployment','Performance falls in an under-represented subgroup'),('Harm','Missed diagnosis, wrong risk estimate or unequal access')]
for i,(h,b) in enumerate(steps):
 x=.65+i*3.16; rect(s,x,2.15,2.7,1.55,[TEAL,BLUE,GOLD,RED][i],True); txt(s,x+.15,2.42,2.4,.32,h,16,WHITE,True,PP_ALIGN.CENTER); txt(s,x+.14,2.88,2.42,.48,b,11,WHITE,False,PP_ALIGN.CENTER)
 if i<3: arrow(s,x+2.72,2.93,x+3.06,2.93,'94BDCF')
card(s,1.15,4.7,5.1,1.2,'Mitigation','Representative data, label audit, subgroup analysis, external validation, monitoring and human escalation pathways.',TEAL)
card(s,7.1,4.7,5.1,1.2,'Key message','Fairness is not guaranteed by a high average AUROC. Report performance in clinically relevant subgroups.',RED)
footer(s,'Bias can arise during sampling, measurement, labelling, model development and use.')
notes(s,'Explain that datasets are not neutral. Representative inclusion, measurement quality and transparent subgroup reporting are needed. A model can be accurate overall while failing the very group that needs it.')
#17
s=new(); title(s,17,'Limitations','Overfitting, generalisability and clinical utility','A model can fit past data well yet fail in new patients or workflows.');
card(s,.9,1.8,3.55,2.15,'Overfitting','Excellent training performance, poor unseen-data performance. Reduce with appropriate complexity control and independent testing.',RED)
card(s,4.9,1.8,3.55,2.15,'Dataset shift','Changes in population, prevalence, scanner, coding or care process after development.',GOLD)
card(s,8.9,1.8,3.55,2.15,'Human factors','Alert fatigue, automation bias, poor interface design and unclear responsibility can reduce safety.',BLUE)
rect(s,.95,4.75,11.35,.8,NAVY,True); txt(s,1.2,4.95,10.85,.35,'Accuracy is only a technical property. Clinical usefulness requires calibration, workflow fit, safety, equity and evidence of patient benefit.',15,WHITE,True,PP_ALIGN.CENTER)
footer(s,'Post-deployment monitoring is needed because performance may drift over time.')
notes(s,'Explain automation bias: users may over-trust a recommendation and ignore contradictory clinical information. The appropriate response is not avoidance of AI, but good design, training, monitoring and accountability.')
#18
s=new(); title(s,18,'Ethics and governance','Human oversight is the safety layer','AI should assist healthcare professionals, not replace clinical responsibility.');
flow=['Patient data','AI analysis','AI-generated insight','Clinician verification','Clinical decision','Patient outcome','Feedback / monitoring']
for i,h in enumerate(flow):
 x=.38+i*1.82; rect(s,x,2.05,1.48,1.03,[BLUE,TEAL,BLUE,GOLD,NAVY,TEAL,RED][i],True); txt(s,x+.06,2.35,1.36,.36,h,11,WHITE,True,PP_ALIGN.CENTER)
 if i<6: arrow(s,x+1.5,2.57,x+1.76,2.57,'8BBBCD')
for i,h in enumerate(['Privacy & consent','Transparency','Accountability','Equity','Safety & security','Human autonomy']): card(s,.65+(i%3)*4.15,4.35+(i//3)*1.0,3.72,.7,h,'',[TEAL,BLUE,GOLD,RED,TEAL,BLUE][i])
footer(s,'WHO (2021): autonomy; well-being/safety; transparency; accountability; inclusiveness/equity; responsive & sustainable AI.')
notes(s,'Use this slide to show where responsibility remains. WHO’s six principles can guide implementation. Key point: a model recommendation must be checked against the patient, clinical context and local standards of care.')
#19
s=new(); title(s,19,'Future','Current, emerging and future possibilities','Separate established applications from plausible but not yet proven possibilities.');
for i,(h,b,c) in enumerate([('Current','Narrow task-specific imaging support, signal analysis, workflow triage, documentation assistance.',TEAL),('Emerging','Multimodal models, generative AI with retrieval, prospective workflow trials, adaptive lifecycle monitoring.',BLUE),('Future possibilities','Digital twins, AI agents and precision diagnostics - require validation, governance and equitable infrastructure.',GOLD)]): card(s,.85+i*4.1,1.92,3.68,2.55,h,b,c)
rect(s,.95,5.25,11.35,.65,'E9F7F7',True); txt(s,1.15,5.43,10.95,.28,'Best future model: human-AI collaboration with transparent evidence, continuous monitoring and patient-centred goals.',14,NAVY,True,PP_ALIGN.CENTER)
footer(s,'Avoid treating speculative applications as established clinical practice.')
notes(s,'Distinguish current deployment from research prototypes. Generative systems are promising for language workflows but must be assessed for factuality, privacy, bias, usability and patient impact before routine high-stakes use.')
#20
s=new(); title(s,20,'Take-home','AI in medicine: evidence before enthusiasm','The most useful question is not “Can AI do this?” but “Does it safely improve care for this population in this setting?”');
for i,t in enumerate(['AI is a tool, not an autonomous replacement for clinical judgment.','Clinical value requires representative data, independent validation and workflow evaluation.','Technical performance must be separated from clinical utility and patient outcomes.','Responsible use requires privacy, equity, transparency, accountability and human oversight.','Medical professionals need AI literacy to use, question and govern these systems.']):
 rect(s,.88,1.72+i*.73,11.6,.52,'E9F7F7' if i%2==0 else WHITE,True); rect(s,1.02,1.88+i*.73,.15,.15,[TEAL,BLUE,GOLD,RED,TEAL][i],True); txt(s,1.35,1.76+i*.73,10.7,.38,t,14,NAVY,True)
txt(s,.9,5.72,11.55,.22,'Thank you  |  Discussion',16,TEAL,True,PP_ALIGN.CENTER)
footer(s,'Selected references on next page in speaker notes. WHO 2021; FDA AI-enabled device list; Harrison’s 22e (2025); PMID 40140300.')
notes(s,'Viva-ready questions and answers:\n1. AI vs ML vs DL? AI is broad; ML learns patterns from data; DL is ML using multi-layer neural networks.\n2. Can AI replace doctors? No. It may support defined tasks, but clinicians retain responsibility for context, communication and decisions.\n3. What is external validation? Testing in a different site, time period or population to assess transportability.\n4. What is hallucination? Fluent generated output that is false or unsupported, including fabricated references.\n5. Why is calibration important? It checks whether predicted risk matches observed risk, essential for risk-based decisions.\n6. Can AI-generated references be trusted? No. Verify every citation in the original source.\n\nReferences (Vancouver style):\n1. World Health Organization. Ethics and governance of artificial intelligence for health. Geneva: WHO; 2021. Available from: https://www.who.int/publications/i/item/9789240029200\n2. World Health Organization. Regulatory considerations on artificial intelligence for health. Geneva: WHO; 2023.\n3. U.S. Food and Drug Administration. Artificial Intelligence-Enabled Medical Devices [Internet]. Silver Spring (MD): FDA; accessed 2026 Sep 19. Available from: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices\n4. Harrison’s Principles of Internal Medicine. 22nd ed. New York: McGraw Hill; 2025. Chapter 501, Machine Learning and Augmented Intelligence in Clinical Medicine.\n5. Aravazhi PS, Gunasekaran PP, Benjamin NZY. The integration of artificial intelligence into clinical medicine: Trends, challenges, and future directions. Dis Mon. 2025;71:101882. doi:10.1016/j.disamonth.2025.101882. PMID:40140300.\n6. Liu X, et al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med. 2020;26:1364-74. doi:10.1038/s41591-020-1034-x.\n7. Collins GS, et al. TRIPOD+AI statement. BMJ. 2024;385:e078378. doi:10.1136/bmj-2023-078378.')

# source note on imaging in speaker notes of relevant slide
notes(prs.slides[5], prs.slides[5].notes_slide.notes_text_frame.text+'\nImage example source in supplied visual assets: medical-image library figure showing AI overlay on chest radiograph; illustrative only, not diagnostic guidance.')
# Save PPT
ppt=OUT+'/AI_in_Medicine_Research_Methodology_Upgraded.pptx'; prs.save(ppt)

# PDF companion, professional readable handout-like slide export
pdf=OUT+'/AI_in_Medicine_Research_Methodology_Upgraded.pdf'; c=canvas.Canvas(pdf,pagesize=landscape((960,540)))
for idx,slide in enumerate(prs.slides,1):
    c.setFillColor(HexColor('#F4F8FB')); c.rect(0,0,960,540,fill=1,stroke=0); c.setFillColor(HexColor('#00A6A6')); c.rect(0,528,960,12,fill=1,stroke=0)
    # extract text from slide, render selective hierarchy
    arr=[]
    for sh in slide.shapes:
        if getattr(sh,'has_text_frame',False):
            t=sh.text.strip()
            if t and not t.isdigit(): arr.append(t.replace('\x0b',' '))
    # title then content, avoid dense overlap
    y=485
    c.setFillColor(HexColor('#0B1F33')); c.setFont('Helvetica-Bold',23); c.drawString(40,y,arr[1] if len(arr)>1 else f'Slide {idx}')
    y-=38; c.setFillColor(HexColor('#5D7180')); c.setFont('Helvetica',10)
    content=' | '.join(arr[2:])
    lines=[]
    for para in content.split('|'):
        lines.extend(textwrap.wrap(para.strip(),110) or [''])
    for line in lines[:24]:
        if y<35: break
        c.drawString(45,y,line); y-=17
    c.setFillColor(HexColor('#00A6A6')); c.setFont('Helvetica-Bold',9); c.drawRightString(920,22,f'{idx:02}')
    c.showPage()
c.save()
print(ppt); print(pdf)

Successfully replaced 1 instance(s) of the string in '/tmp/workspace/ai-medicine-presentation/src/build_deck.py'