Zyadatar Clinical AI Validation Studies Misleading Kyun Hain

Jab koi clinical AI company validation results publish karti hai, toh headline metric lagbhag hamesha AUROC hota hai: Area Under the Receiver Operating Characteristic curve. 0.95 ka AUROC kamaal ka lagta hai, aur ye pitch decks, press releases aur conference talks mein baar baar is saboot ki tarah aata hai ki system "kaam karta hai." Hospital administrators, investors, aur kuch clinicians bhi 0.95 dekhte hain aur maan lete hain ki system 95% baar target condition ko sahi pehchaanta hai.

Ye matlab nikaalna galat hai.

AUROC discrimination naapta hai: system kitni achhi tarah ek randomly chune gaye positive case ko ek randomly chune gaye negative case se upar rank karta hai. Ye saare mumkin classification thresholds ka ek summary hai, jinmein se zyadatar practice mein kabhi use nahi honge. Baaki sab barabar ho toh 0.95 AUROC wala system 0.80 wale se ranking mein behtar hai. Lekin AUROC ye kuch nahi batata ki us khaas operating threshold pe kya hota hai jo hospital asal mein deploy karega, aur ye bhi nahi ki system us population mein kaisa perform karta hai jise hospital asal mein serve karta hai.

Ek sepsis detection system lijiye: AUROC 0.95, 85% sensitivity, aur 90% specificity. Paper mein ye numbers strong lagte hain. Ab inhe ek asli ICU mein rakhiye jahan sepsis ki prevalence 5% hai.

Har 1,000 screened patients mein: 50 ko sepsis hai, 950 ko nahi. System 50 sepsis patients mein se 42.5 ko sahi pehchaanta hai. Bina sepsis wale 950 mein se 855 ko sahi clear karta hai. Lekin unhi 950 non-sepsis patients se wo 95 false alarms bhi deta hai. Positive predictive value, yaani flag hue patient ko sach mein sepsis hone ki probability, 42.5 / (42.5 + 95) = 30.9% hai.

Is system ko use karne wala clinician har ek sahi alarm pe lagbhag do false alarms dekhega. Ek busy ICU mein jo system alert dene pe 70% baar galat ho, use kuch hi hafton mein noise maan liya jaayega. 0.95 ka AUROC asli hai, lekin wo us false positive bojh ko chhupa deta hai jo tay karta hai ki koi system ko asal mein use karega ya nahi.

Maine ye hote hue clinical training ke dauraan dekha, AI mein kaam karne se bahut pehle. EHR se alert fatigue tab bhi itni zyada thi ki physicians ke haath alert ka text padhe bina "override" click karne ke aadi ho chuke the.

Zyadatar clinical AI validation studies system ko usi institution ke data pe test karti hain jisne use develop kiya. Wajah practical hai: developers ke paas apna data hota hai, wo apne documentation patterns samajhte hain, aur internal validation tez aur sasta hota hai. Lekin isse ek systematic bias paida hota hai jo reported performance ko badha chadha deta hai.

Hospitals aise tareekon se alag hote hain jo algorithm ki performance pe gehra asar daalte hain. Documentation practices alag hoti hain: kuch institutions vital signs har ghante chart karte hain, kuch har chaar ghante. Lab order karne ke patterns alag hain. EHR configurations standardized nahi hain. Jab ek model ek hi institution ke data pe train aur test hota hai, toh wo aise patterns seekhta hai jinmein asli clinical signals ke saath ye saare site-specific artifacts bhi shaamil hote hain. Internal validation dono ko pakadta hai. External validation dikhata hai ki kaun se signals asli hain.

Sabse seekhne layak example Epic ka sepsis prediction model hai. Internally train aur validate hone pe usne strong discrimination dikhaya. Jab University of Michigan ke researchers ne usi model ko apne patients pe independently evaluate kiya, toh AUROC 0.63 pe gir gaya, jo chance se bas zara sa behtar hai. Wong et al. ki ye study, jo 2021 mein JAMA Internal Medicine mein publish hui, mein paaya gaya ki model ne sepsis ke do-tihai cases miss kiye aur itne false alerts diye ki clinicians ka bharosa tezi se toot gaya.

Korean Journal of Radiology mein ek systematic review ne diagnostic imaging AI ki 516 published studies dekhin aur paaya ki sirf 6% ne apne algorithm ko ek independent, external dataset pe test kiya. Jo hospitals baaki 94% ke basis pe kharidne ka faisla karte hain, wo ek tarah se aise performance claims khareed rahe hain jo verify hi nahi hue.

Ek company bilkul sahi numbers report karke bhi galat tasveer bana sakti hai. Ye fraud ki baat nahi hai. Ye framing ki baat hai.

Chuninda subgroup reporting: jis system ka overall AUROC 0.82 hai, wo shayad 65 se upar ke aur kam se kam teen comorbidities wale patients mein 0.91 tak pahunch jaaye. Company "high-risk population" ke liye 0.91 report karti hai, jo technically sahi hai, lekin aisa impression deta hai ki system poori patient population mein jitna achha hai, usse behtar hai.

PPV ke bina sensitivity aur specificity report karna: jo company kehti hai "92% sensitivity aur 88% specificity", wo sach bol rahi hai aur numbers impressive lagte hain. Jo company kehti hai "4% prevalence pe 28% PPV", wo usi system ki baat kar rahi hai. Sales presentation mein aap kaun sa number pehle rakhenge?

Ek zimmedaar validation report us threshold pe performance dikhati hai jo deployment mein asal mein use hota hai, sirf saare mumkin thresholds ka nahi. Wo dikhati hai ki system alag tarah ke hospitals aur patient populations mein kaisa perform karta hai, sirf ek pooled number nahi. Aur wo batati hai ki validation internal tha ya external. Zyadatar reports aisa nahi karti.

Clinical AI evaluate kar rahe zyadatar health systems ko ek slide deck dikhaya jaata hai jismein ek hi performance number hota hai. Wo number aam taur pe AUROC hota hai, ek aisa metric jise samajhne ki training zyadatar clinicians aur hospital administrators ko kabhi mili hi nahi. Ye naapta hai ki system patients ko risk ke hisaab se kitni achhi tarah rank karta hai, ye nahi ki alert dene pe wo kitna sahi hai. High AUROC ka matlab ye nahi ki system aapke hospital mein, aapke patients ke saath, us threshold pe jo aap use karenge, achha kaam karega.

Zyada matter ye karta hai. Kya system ko un hospitals mein test kiya gaya jo use banane mein shaamil nahi the? Agar sirf internally validate hua hai, toh ho sakta hai performance aapke institution mein na tike. Alag hospitals alag tarah se document karte hain, alag frequency pe labs order karte hain, aur alag populations ko serve karte hain. Jo system us hospital mein kaam karta hai jahan wo bana, usne ye bahut kam saabit kiya hai ki wo aur kahan kaam karega.

Aapki nurses ko har shift mein kitne false alarms dikhenge? Agar system har unit pe 20 alerts deta hai aur unmein se 15 galat hain, toh aapka staff kuch hi hafton mein sab ko ignore karna seekh jaayega. Us asli threshold pe false positive rate maangiye jo system use karta hai, validation paper ka theoretical best case nahi.

Wo problems asal mein kitni jaldi detect karta hai? Jo system crisis se 2 ghante pehle deterioration pakadta hai, uski kuch value hai. Jo ek din ya usse zyada pehle pakadta hai, wo aapki team ko bilkul alag interventions ka time deta hai. Asli deployments se median lead time maangiye, retrospective analysis se nahi.

Kya aapke clinicians dekh sakte hain ki system ne kisi patient ko flag kyun kiya? Agar jawaab nahi hai, toh accuracy chahe jo ho, adoption kam hi rahega. Physicians ko evidence parakhne ki training milti hai, bina wajah ke scores pe act karne ki nahi.

ML community ne discrimination metrics ke liye lagataar optimize kiya hai. AUROC aur AUPRC is sawaal ka jawaab dete hain: kya system positive aur negative cases ko alag kar sakta hai? Clinical deployment ke liye ye zaroori hai, lekin kaafi nahi.

Calibration naapta hai ki predicted probabilities absolute terms mein bharose layak hain ya nahi. Ek well-calibrated system jo sepsis ka 30% risk deta hai, us group mein lagbhag 30% ko sach mein sepsis hona chahiye. Kharab calibration ka matlab hai ki numbers pe jaise ke taise bharosa nahi kiya ja sakta, aur jo clinician watchful waiting aur aggressive intervention ke beech faisla kar raha hai, use pata hona chahiye ki "70% risk" ka matlab sach mein 70% hai ya nahi. Calibration ko calibration curves, Brier scores aur Hosmer-Lemeshow statistic se assess kiya ja sakta hai. Clinical AI publications mein ye shayad hi kabhi report hota hai. Is kami ko ek warning sign maanna chahiye.

Lead time naapta hai ki conventional approaches ke muqable system conditions ko kitna pehle detect karta hai. Jo systems "early warning" dene ka daava karte hain, unke liye asli sawaal sirf ye nahi ki wo deterioration pehchaan sakte hain ya nahi, balki ye ki standard clinical assessment ke muqable wo use kitna pehle pehchaante hain. Kam AUROC lekin 20 ghante zyada lead time wala system, high AUROC lekin na ke barabar early detection wale system se zyada jaanein bacha sakta hai.

Hum khud ko bhi inhi standards pe parakhte hain. Hamara apna validation, calibration aur lead time samet, poora hone pe publish kiya jaayega.